[llvm-branch-commits] [llvm] 7273822 - Revert "[AMDGPU] add tests for loop definition of bitconvert (#133052)"
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Apr 7 23:03:15 PDT 2025
Author: Shoreshen
Date: 2025-04-08T14:02:43+08:00
New Revision: 7273822d844c78294e1e60fc9c901df538579063
URL: https://github.com/llvm/llvm-project/commit/7273822d844c78294e1e60fc9c901df538579063
DIFF: https://github.com/llvm/llvm-project/commit/7273822d844c78294e1e60fc9c901df538579063.diff
LOG: Revert "[AMDGPU] add tests for loop definition of bitconvert (#133052)"
This reverts commit 49d6e39eba88fdf7975c6f4dcb38595538331e6a.
Added:
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.ll
Modified:
Removed:
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.160bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.192bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.224bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.288bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.352bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.384bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.448bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.ptr.ll
################################################################################
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
deleted file mode 100644
index 51dc5ceb82b41..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ /dev/null
@@ -1,108541 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <32 x float> @bitcast_v32i32_to_v32f32(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <32 x i32> @bitcast_v32f32_to_v32i32(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB1_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB1_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <16 x i64> @bitcast_v32i32_to_v16i64(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <32 x i32> @bitcast_v16i64_to_v32i32(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB3_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB3_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB3_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <16 x double> @bitcast_v32i32_to_v16f64(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB4_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB4_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB4_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <32 x i32> @bitcast_v16f64_to_v32i32(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GCN-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GCN-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GCN-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GCN-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GCN-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GCN-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GCN-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB6_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 3, v58
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 3, v59
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB6_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v52
-; GCN-NEXT: v_or_b32_e32 v1, v1, v52
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_or_b32_e32 v2, v2, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v50, v31
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v31
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v50
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v49
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v61
-; GCN-NEXT: v_or_b32_e32 v49, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v48
-; GCN-NEXT: v_or_b32_e32 v2, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v60
-; GCN-NEXT: v_or_b32_e32 v61, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v39
-; GCN-NEXT: v_or_b32_e32 v62, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v57
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v38
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v56
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v37
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v47
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v36
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v46
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v35
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v45
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v34
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v44
-; GCN-NEXT: v_or_b32_e32 v14, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v33
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v43
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v21
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v42
-; GCN-NEXT: v_or_b32_e32 v18, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v41
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v25
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v40
-; GCN-NEXT: v_or_b32_e32 v22, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v27
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v55
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v25, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v26, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v27, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v53
-; GCN-NEXT: v_or_b32_e32 v28, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v29, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v30, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v1, v29
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v49
-; GCN-NEXT: v_or_b32_e32 v5, v1, v32
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v30
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v62
-; GCN-NEXT: v_or_b32_e32 v6, v6, v34
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v36
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v37
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v38
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v48
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v9, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v11, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v12, v12, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v13, v13, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v15, v15, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v17, v17, v42
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v43
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v19, v44
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v45
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v21, v21, v46
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v47
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v23, v23, v56
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v57
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v25, v25, v58
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v59
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v27, v27, v60
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB6_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v39, 24, v32
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_mov_b32_e32 v55, v39
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: .LBB6_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB6_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v32, vcc, 3, v32
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: .LBB6_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v48
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v57
-; VI-NEXT: v_or_b32_sdwa v2, v2, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v41
-; VI-NEXT: v_or_b32_sdwa v48, v53, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v38
-; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v47
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v45
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v46, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v37
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v49
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB6_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_waitcnt vmcnt(29)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(31)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: .LBB6_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB6_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(29)
-; GFX9-NEXT: v_add_u32_e32 v32, 3, v32
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: .LBB6_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v52
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v56
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v47
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v46
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v41
-; GFX9-NEXT: v_or_b32_sdwa v33, v44, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_or_b32_sdwa v1, v53, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB6_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: .LBB6_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB6_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_add_nc_u32_e32 v32, 3, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: .LBB6_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v63
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v65
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v61
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v73
-; GFX11-NEXT: v_or_b32_e32 v65, v67, v65
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v66
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v62
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v66
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v60
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v65
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v67
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v56
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v65
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v47
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v45
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v64
-; GFX11-NEXT: v_or_b32_e32 v55, v65, v66
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v44
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v43
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v54, v65, v54
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v40
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v55
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v65
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v182
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v181
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v55
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v65
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v54
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v53, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v176
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v167
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v53, v53, v54
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v52
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v65
-; GFX11-NEXT: v_or_b32_e32 v54, v66, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v53
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v54
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v164
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v160
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v150
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v146
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v49
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v145
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v144
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v135
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v132
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v131
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v128
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v119
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v117
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v116
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v115
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v103
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v102
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v101
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v99
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v86
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v84
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v83
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v81
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v70
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v69
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v68
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v10
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:148
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:160
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:192
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:212
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:244
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:288
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:340
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:352
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:380
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:376
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v1
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:372
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:364
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v0, v39
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v47
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v54
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v46
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v37
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v48
-; GCN-NEXT: v_or_b32_e32 v8, v8, v45
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v53
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v42
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v41
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v63
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v50
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v49
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v20, v60
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v58
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v22, v62
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v59
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v61
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v52
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v35
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v33
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v44
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v57
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v38, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v48, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v55, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v54, v49
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v18, v18, v50
-; GCN-NEXT: v_or_b32_e32 v19, v19, v51
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_or_b32_e32 v21, v21, v53
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: v_or_b32_e32 v23, v23, v33
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_or_b32_e32 v25, v25, v35
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v37
-; GCN-NEXT: v_or_b32_e32 v28, v28, v38
-; GCN-NEXT: v_or_b32_e32 v29, v29, v39
-; GCN-NEXT: v_or_b32_e32 v30, v30, v48
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; kill: killed $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB7_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v39, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v56, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v47, v3
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v54, v4
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v46, v5
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v37, v6
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v43, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v45, v8
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v42, v10
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v41, v11
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v40, v12
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v63, v13
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v50, v14
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v0, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v0, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v0, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v51, v18
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v49, v19
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v60, v20
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v58, v21
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v25, v62, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v29, v59, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v37, v32, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v50, v61, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v41, v34, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v45, v52, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v56, v35, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v58, v33, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v59, v44, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v57
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v57, v36, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v26, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v48, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v53, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_mov_b32_e32 v0, v55
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v55, v53
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v40, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v40
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v42, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v44, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v46, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v46
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_or_b32_e32 v46, v0, v46
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v47
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v0, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v60, v0
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v61, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v62, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v63, v3
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s7, v37
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s7, v50
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s7, v41
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s7, v45
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s7, v56
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 0x300, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff, v56
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_or_b32_e32 v4, v36, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v39, v6
-; GCN-NEXT: v_or_b32_e32 v7, v49, v7
-; GCN-NEXT: v_or_b32_e32 v8, v51, v8
-; GCN-NEXT: v_or_b32_e32 v9, v52, v9
-; GCN-NEXT: v_or_b32_e32 v10, v54, v10
-; GCN-NEXT: v_or_b32_e32 v11, v22, v11
-; GCN-NEXT: v_or_b32_e32 v12, v23, v12
-; GCN-NEXT: v_or_b32_e32 v13, v24, v13
-; GCN-NEXT: v_or_b32_e32 v14, v26, v14
-; GCN-NEXT: v_or_b32_e32 v15, v27, v15
-; GCN-NEXT: v_or_b32_e32 v16, v28, v16
-; GCN-NEXT: v_or_b32_e32 v17, v30, v17
-; GCN-NEXT: v_or_b32_e32 v18, v31, v18
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: v_or_b32_e32 v20, v33, v20
-; GCN-NEXT: v_or_b32_e32 v21, v34, v21
-; GCN-NEXT: v_or_b32_e32 v22, v35, v25
-; GCN-NEXT: v_or_b32_e32 v23, v48, v29
-; GCN-NEXT: v_or_b32_e32 v24, v53, v37
-; GCN-NEXT: v_or_b32_e32 v25, v55, v50
-; GCN-NEXT: v_or_b32_e32 v26, v40, v41
-; GCN-NEXT: v_or_b32_e32 v27, v42, v45
-; GCN-NEXT: v_or_b32_e32 v28, v43, v56
-; GCN-NEXT: v_or_b32_e32 v29, v44, v58
-; GCN-NEXT: v_or_b32_e32 v30, v46, v59
-; GCN-NEXT: v_or_b32_e32 v31, v47, v57
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x3000000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x3000000, v31
-; GCN-NEXT: .LBB7_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: .LBB7_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v31, 0x300
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v61
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v63
-; VI-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v59
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v62
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v58
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v60
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v54
-; VI-NEXT: v_add_u16_e32 v15, 3, v40
-; VI-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v14, v15
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v15, v16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v16, v16, v17
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v18
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v18, 0x300, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v18, v19
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v19, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 0x300, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v20, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v21, 0x300, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v21, v22
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v22, v23
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v23, v24
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v24, v25
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v25, v26
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v26, v27
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v27, v28
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v28, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v32, 0x300, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v33, 3, v33
-; VI-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: .LBB7_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: .LBB7_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v54
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v29, v29, v30
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v30, v30, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v31, v31, v32
-; GFX9-NEXT: .LBB7_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:592
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:588
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:584
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:468
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:456
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v136, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v137, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v138, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
-; GFX11-NEXT: v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
-; GFX11-NEXT: v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
-; GFX11-NEXT: v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
-; GFX11-NEXT: v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
-; GFX11-NEXT: v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
-; GFX11-NEXT: v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v114, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v115, off, s32
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v181, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v182, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v136, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v137, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v138, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v132, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v150, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v160, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v161, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v167, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v176, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v41, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v115
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v116
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v117
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v118
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v119
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v131
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v144
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v145
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v146
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v147
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v148
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v162
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v163
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v164
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v165
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v166
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v179, 8, v179
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v180, 8, v180
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v181
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v183, 8, v183
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v162, 8, v136
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v163, 8, v137
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v164, 8, v138
-; GFX11-NEXT: v_lshlrev_b16 v165, 8, v103
-; GFX11-NEXT: v_lshlrev_b16 v166, 8, v102
-; GFX11-NEXT: v_lshlrev_b16 v144, 8, v101
-; GFX11-NEXT: v_lshlrev_b16 v145, 8, v100
-; GFX11-NEXT: v_lshlrev_b16 v146, 8, v99
-; GFX11-NEXT: v_lshlrev_b16 v147, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v148, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v28
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v26
-; GFX11-NEXT: v_lshlrev_b16 v129, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v130, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v131, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v14
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v0
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v51
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v124
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v125
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v126
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v127
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v111
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v121
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v120
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v122
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v123
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v107
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v108
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v109
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v110
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v106
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v7, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v6, v12
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v92
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v78
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v77
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v76
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v75
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v59
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v93
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v94
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v95
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v104
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v105
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v79
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v88
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v89
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v90
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v91
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v58
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v44
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v43
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v41
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v40
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v176
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v167
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v61
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v62
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v63
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v72
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v73
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v45
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v46
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v47
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v56
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v57
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v161
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v160
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v150
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v132
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v113
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v179
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v180
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v181
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v182
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v183
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v162
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v163
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v164
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v165
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v166
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v84
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v82
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v144
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v145
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v146
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v147
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v148
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v119
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v128
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v129
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v130
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v131
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v66
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v64
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v114
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v115
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v116
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v117
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v118
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v99
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v100
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v101
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v102
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v103
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: .LBB7_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v51, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v5, v50, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v124, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v125, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v126, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v127, v3
-; GFX11-NEXT: v_add_nc_u16 v6, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v7, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_add_nc_u16 v8, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v36, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v10, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v2, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v11, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_or_b32_e32 v3, v111, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v120, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v121, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v6, v122, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v107, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v108, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v109, v10
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v38, 3
-; GFX11-NEXT: v_or_b32_e32 v11, v110, v11
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v106, v2
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_add_nc_u16 v7, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v32, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v9, v92, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v78, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v77, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v76, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v75, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v74, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v60, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v59, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v93, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v94, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v95, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v104, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v105, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v79, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v88, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v89, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v90, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v91, v16
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_add_nc_u16 v12, v58, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v44, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v43, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v41, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v40, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v178, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v177, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v176, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v167, 3
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v61, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v62, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v63, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v72, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v73, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v45, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v46, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v47, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v56, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v57, v21
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_add_nc_u16 v17, v161, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v160, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v150, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v149, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v133, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v132, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v113, 3
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v179, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v180, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v181, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v182, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v183, v21
-; GFX11-NEXT: v_or_b32_e32 v22, v162, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v163, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v164, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v165, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v166, v26
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_add_nc_u16 v22, v112, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v97, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v87, 3
-; GFX11-NEXT: v_add_nc_u16 v27, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v84, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v83, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v82, 3
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v144, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v145, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v146, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v147, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v148, v26
-; GFX11-NEXT: v_or_b32_e32 v27, v119, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v128, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v129, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v130, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v131, v31
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_add_nc_u16 v27, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v33, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v34, v66, 3
-; GFX11-NEXT: v_add_nc_u16 v35, v65, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v64, 3
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v114, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v115, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v116, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v117, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v118, v31
-; GFX11-NEXT: v_or_b32_e32 v32, v99, v32
-; GFX11-NEXT: v_or_b32_e32 v33, v100, v33
-; GFX11-NEXT: v_or_b32_e32 v34, v101, v34
-; GFX11-NEXT: v_or_b32_e32 v35, v102, v35
-; GFX11-NEXT: v_or_b32_e32 v36, v103, v36
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v34
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v36
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: .LBB7_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v138, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v137, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v136, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:516
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:580
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:584
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:588
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:592
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v62
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: .LBB8_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v62
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: .LBB8_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v61
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v59
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v56, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v47
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v46, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v45, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v44, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 20, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v33
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v17, 16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v18, v18, v19, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v20, v20, v21, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_alignbit_b32 v24, v24, v25, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_alignbit_b32 v26, v26, v27, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_alignbit_b32 v28, v28, v29, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_alignbit_b32 v30, v30, v31, 16
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_alignbit_b32 v32, v32, v33, 16
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_alignbit_b32 v34, v34, v35, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_alignbit_b32 v36, v36, v37, 16
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v38, v38, v39, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_alignbit_b32 v48, v48, v49, 16
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_alignbit_b32 v50, v50, v51, 16
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_alignbit_b32 v52, v52, v53, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_alignbit_b32 v54, v54, v55, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_alignbit_b32 v40, v40, v41, 16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v63
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_alignbit_b32 v42, v42, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v56, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB8_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: .LBB8_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB8_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: .LBB8_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB8_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: .LBB8_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <32 x i32> @bitcast_v64bf16_to_v32i32(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v33
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v42
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v51
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v63
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v60
-; GCN-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v61, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v58
-; GCN-NEXT: v_alignbit_b32 v2, v2, v59, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v56
-; GCN-NEXT: v_alignbit_b32 v3, v3, v57, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v46
-; GCN-NEXT: v_alignbit_b32 v4, v4, v47, 16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v45, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v11, v11, v12, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v13, v13, v14, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v34
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v33, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v17, v18, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v43
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v19, v19, v42, 16
-; GCN-NEXT: v_alignbit_b32 v20, v20, v44, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; GCN-NEXT: v_alignbit_b32 v22, v22, v48, 16
-; GCN-NEXT: v_alignbit_b32 v23, v23, v38, 16
-; GCN-NEXT: v_alignbit_b32 v24, v24, v50, 16
-; GCN-NEXT: v_alignbit_b32 v25, v25, v51, 16
-; GCN-NEXT: v_alignbit_b32 v26, v26, v53, 16
-; GCN-NEXT: v_alignbit_b32 v27, v27, v55, 16
-; GCN-NEXT: v_alignbit_b32 v28, v28, v40, 16
-; GCN-NEXT: v_alignbit_b32 v29, v29, v63, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; kill: killed $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: .LBB9_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v4, v5, v4, 16
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_alignbit_b32 v6, v7, v6, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v7, v8, v7, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v8, v9, v8, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_alignbit_b32 v10, v11, v10, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v11, v12, v11, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v12, v13, v12, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_alignbit_b32 v14, v15, v14, 16
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v34
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v15, v16, v15, 16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v33
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_alignbit_b32 v16, v17, v16, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v42
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v44
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v54
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v41
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v43
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v40, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v44, 0x40c00000, v31
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v32
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v33
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v18, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v19, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v20, v39, v54, 16
-; GCN-NEXT: v_alignbit_b32 v21, v48, v40, 16
-; GCN-NEXT: v_alignbit_b32 v22, v49, v22, 16
-; GCN-NEXT: v_alignbit_b32 v23, v50, v23, 16
-; GCN-NEXT: v_alignbit_b32 v24, v51, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v32, v25, 16
-; GCN-NEXT: v_alignbit_b32 v26, v33, v26, 16
-; GCN-NEXT: v_alignbit_b32 v27, v34, v27, 16
-; GCN-NEXT: v_alignbit_b32 v28, v35, v28, 16
-; GCN-NEXT: v_alignbit_b32 v29, v36, v29, 16
-; GCN-NEXT: v_alignbit_b32 v30, v37, v30, 16
-; GCN-NEXT: v_alignbit_b32 v31, v38, v31, 16
-; GCN-NEXT: .LBB9_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v15
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v14
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v13
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v12
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v11
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v10
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v9
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v8
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v7
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v6
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v5
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v4
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v3
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v2
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v1
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v0
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v31
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v30
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v29
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v29, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v28
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v28, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v27
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v27, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v26
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v26, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v25
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v25, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v24
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v24, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v23
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v23, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v22
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v22, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v21
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v20
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v20, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v19
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v19, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v18
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v17
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v17, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB9_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v32, s7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v31, v31, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v30, v30, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v29, v29, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v28, v28, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v27, v27, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v26, v26, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v25, v25, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v24, v24, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v23, v23, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v22, v22, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v21, v21, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v20, v20, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v19, v19, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v18, v18, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v17, v17, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v16, v16, v32, s7
-; GFX9-NEXT: .LBB9_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v32, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v32
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v39, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_add3_u32 v34, v34, v32, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v39, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v36, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v36, v36, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v32, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_perm_b32 v14, v14, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v33, v33, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v37, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v13, v13, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v34, v34, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v33, v33, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_bfe_u32 v35, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v35, v35, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v33, v33, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v35, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v33, v33, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GFX11-NEXT: v_perm_b32 v7, v7, v32, 0x7060302
-; GFX11-NEXT: v_perm_b32 v6, v6, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_bfe_u32 v34, v4, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v5, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v3
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v4
-; GFX11-NEXT: v_perm_b32 v5, v5, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v4, v4, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v35, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v2, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_perm_b32 v3, v3, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_perm_b32 v2, v2, v33, 0x7060302
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v34, v0, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v0
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_perm_b32 v1, v1, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v0, v0, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX11-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_bfe_u32 v35, v30, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v29
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v30, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_perm_b32 v31, v31, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v30, v30, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_add3_u32 v33, v33, v29, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v29
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-NEXT: v_bfe_u32 v34, v28, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v27
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v28, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v28
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v26
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_perm_b32 v29, v29, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v28, v28, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX11-NEXT: v_add3_u32 v33, v33, v27, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v27
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-NEXT: v_bfe_u32 v35, v26, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v25
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v26, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v26
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_perm_b32 v27, v27, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v26, v26, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-NEXT: v_add3_u32 v33, v33, v25, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v25
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-NEXT: v_bfe_u32 v34, v24, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v23
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v24
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_perm_b32 v25, v25, v32, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v24, v24, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v33, v23, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-NEXT: v_bfe_u32 v35, v22, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v22, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v22
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v23, v23, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v36, v34, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v21, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v21
-; GFX11-NEXT: v_perm_b32 v22, v22, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GFX11-NEXT: v_add3_u32 v35, v36, v34, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v21, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v33, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v32, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v32, v38, v33, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v21, v21, v34, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v34, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v35, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_add3_u32 v34, v34, v20, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v34, v35, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v36
-; GFX11-NEXT: v_add3_u32 v35, v37, v33, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v33
-; GFX11-NEXT: v_bfe_u32 v37, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_perm_b32 v20, v20, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v37, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v37, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_bfe_u32 v48, v18, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v36, v48, v18, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX11-NEXT: v_perm_b32 v19, v19, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v36, v37, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-NEXT: v_add3_u32 v37, v38, v35, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v35
-; GFX11-NEXT: v_bfe_u32 v39, v17, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_bfe_u32 v48, v36, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v36
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v16
-; GFX11-NEXT: v_perm_b32 v18, v18, v34, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v39, v17, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_bfe_u32 v37, v16, 16, 1
-; GFX11-NEXT: v_add3_u32 v48, v48, v36, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v38, v39, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_add3_u32 v37, v37, v16, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v17, v17, v35, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v48, v49, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v37, v50, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v16, v16, v36, 0x7060302
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <64 x half> @bitcast_v32i32_to_v64f16(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v62
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v62
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v63
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v36
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v38
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v48
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v49
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v50
-; GCN-NEXT: v_mov_b32_e32 v50, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB10_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v31, v33
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v61
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: .LBB10_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v59
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v58
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v2, v1
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v57
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v2, v1
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 12, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v47
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v2, v1
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v46
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 20, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v45
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v44
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v43
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v42
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v41
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v54
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v53
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v49
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v48
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v36
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v49, v48
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v52, v53, v52
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v55, v54
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v40, v41, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v61, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v60, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v46, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB10_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: .LBB10_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB10_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: .LBB10_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB10_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: .LBB10_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <32 x i32> @bitcast_v64f16_to_v32i32(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v43
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v50
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v1
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v61
-; GCN-NEXT: v_or_b32_e32 v0, v62, v0
-; GCN-NEXT: v_or_b32_e32 v1, v60, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: v_or_b32_e32 v2, v58, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v57
-; GCN-NEXT: v_or_b32_e32 v3, v56, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v47
-; GCN-NEXT: v_or_b32_e32 v4, v46, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v32
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v37
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_or_b32_e32 v19, v43, v19
-; GCN-NEXT: v_or_b32_e32 v20, v41, v20
-; GCN-NEXT: v_or_b32_e32 v21, v55, v21
-; GCN-NEXT: v_or_b32_e32 v22, v49, v22
-; GCN-NEXT: v_or_b32_e32 v23, v50, v23
-; GCN-NEXT: v_or_b32_e32 v24, v39, v24
-; GCN-NEXT: v_or_b32_e32 v25, v36, v25
-; GCN-NEXT: v_or_b32_e32 v26, v48, v26
-; GCN-NEXT: v_or_b32_e32 v27, v52, v27
-; GCN-NEXT: v_or_b32_e32 v28, v53, v28
-; GCN-NEXT: v_or_b32_e32 v29, v54, v29
-; GCN-NEXT: v_or_b32_e32 v30, v40, v30
-; GCN-NEXT: v_or_b32_e32 v31, v42, v31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: .LBB11_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v41
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v55
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v50
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v39
-; GCN-NEXT: v_mov_b32_e32 v39, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v48
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v33
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_or_b32_e32 v19, v21, v20
-; GCN-NEXT: v_or_b32_e32 v20, v55, v39
-; GCN-NEXT: v_or_b32_e32 v21, v41, v48
-; GCN-NEXT: v_or_b32_e32 v22, v22, v49
-; GCN-NEXT: v_or_b32_e32 v23, v23, v50
-; GCN-NEXT: v_or_b32_e32 v24, v24, v51
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v38
-; GCN-NEXT: v_or_b32_e32 v28, v28, v33
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_or_b32_e32 v30, v30, v35
-; GCN-NEXT: v_or_b32_e32 v31, v31, v37
-; GCN-NEXT: .LBB11_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <64 x i16> @bitcast_v32i32_to_v64i16(<32 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i32_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; kill: killed $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; kill: killed $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v50, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v52, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v55, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v41, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB12_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v50, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v52, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v55, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v41, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB12_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v59
-; GCN-NEXT: v_or_b32_e32 v1, v1, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v33
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v57
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v46
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v62
-; GCN-NEXT: v_or_b32_e32 v46, v1, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v43
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v41
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v60
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v55
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v58
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v52
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v56
-; GCN-NEXT: v_or_b32_e32 v14, v14, v33
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v15, v15, v50
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v47
-; GCN-NEXT: v_or_b32_e32 v16, v16, v52
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v17, v17, v48
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v45
-; GCN-NEXT: v_or_b32_e32 v18, v18, v55
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_or_b32_e32 v19, v19, v39
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v20, v20, v34
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v38
-; GCN-NEXT: v_or_b32_e32 v21, v21, v34
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v42
-; GCN-NEXT: v_or_b32_e32 v22, v22, v34
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v37
-; GCN-NEXT: v_or_b32_e32 v23, v23, v34
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v40
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v36
-; GCN-NEXT: v_or_b32_e32 v25, v25, v34
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v54
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v35
-; GCN-NEXT: v_or_b32_e32 v27, v27, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v53
-; GCN-NEXT: v_or_b32_e32 v28, v28, v34
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_or_b32_e32 v30, v30, v51
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_or_b32_e32 v32, v32, v43
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v44, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i32_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: .LBB12_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i32_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v31, 3, v31
-; GFX9-NEXT: v_add_u32_e32 v30, 3, v30
-; GFX9-NEXT: v_add_u32_e32 v29, 3, v29
-; GFX9-NEXT: v_add_u32_e32 v28, 3, v28
-; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: .LBB12_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i32_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u32_e32 v31, 3, v31
-; GFX11-NEXT: v_add_nc_u32_e32 v30, 3, v30
-; GFX11-NEXT: v_add_nc_u32_e32 v29, 3, v29
-; GFX11-NEXT: v_add_nc_u32_e32 v28, 3, v28
-; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
-; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
-; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
-; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
-; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
-; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
-; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
-; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: .LBB12_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i32> %a, splat (i32 3)
- %a2 = bitcast <32 x i32> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i32> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v32i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v20
-; GCN-NEXT: v_mov_b32_e32 v38, v18
-; GCN-NEXT: v_mov_b32_e32 v39, v16
-; GCN-NEXT: v_mov_b32_e32 v48, v14
-; GCN-NEXT: v_mov_b32_e32 v49, v12
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v51, v8
-; GCN-NEXT: v_mov_b32_e32 v52, v6
-; GCN-NEXT: v_mov_b32_e32 v53, v4
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(12) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v24
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v55
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v36
-; GCN-NEXT: v_or_b32_e32 v1, v1, v58
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v53
-; GCN-NEXT: v_or_b32_e32 v2, v2, v57
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v52
-; GCN-NEXT: v_or_b32_e32 v3, v3, v35
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v4, v4, v60
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v50
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v49
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v48
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v39
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v38
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v43
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v46
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v40
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v62
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v47
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v44
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v18, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v21, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v29, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v30, v32
-; GCN-NEXT: v_or_b32_e32 v31, v31, v59
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: .LBB13_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v36, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v58, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v57, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v52
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v35, v3
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v60, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v32, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v32, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v15
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v32, v16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v32, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v32, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v32, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v32, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v32, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v32, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v32, v30
-; GCN-NEXT: v_or_b32_e32 v31, v59, v31
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x30000, v31
-; GCN-NEXT: .LBB13_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(11)
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v32i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v33, 3
-; VI-NEXT: v_add_u16_e32 v32, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v32, v15
-; VI-NEXT: v_add_u16_e32 v32, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v32, v14
-; VI-NEXT: v_add_u16_e32 v32, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v32, v13
-; VI-NEXT: v_add_u16_e32 v32, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v32, v12
-; VI-NEXT: v_add_u16_e32 v32, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v32, v11
-; VI-NEXT: v_add_u16_e32 v32, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v32, v10
-; VI-NEXT: v_add_u16_e32 v32, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v32, v9
-; VI-NEXT: v_add_u16_e32 v32, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v32, v8
-; VI-NEXT: v_add_u16_e32 v32, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v32, v7
-; VI-NEXT: v_add_u16_e32 v32, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v32, v6
-; VI-NEXT: v_add_u16_e32 v32, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v32, v5
-; VI-NEXT: v_add_u16_e32 v32, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v32, v4
-; VI-NEXT: v_add_u16_e32 v32, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v32, v3
-; VI-NEXT: v_add_u16_e32 v32, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v32, v2
-; VI-NEXT: v_add_u16_e32 v32, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v32, v1
-; VI-NEXT: v_add_u16_e32 v32, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v32, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v31
-; VI-NEXT: v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: v_add_u16_e32 v32, 3, v30
-; VI-NEXT: v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v32, v30
-; VI-NEXT: v_add_u16_e32 v32, 3, v29
-; VI-NEXT: v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v32, v29
-; VI-NEXT: v_add_u16_e32 v32, 3, v28
-; VI-NEXT: v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v32, v28
-; VI-NEXT: v_add_u16_e32 v32, 3, v27
-; VI-NEXT: v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_add_u16_e32 v32, 3, v26
-; VI-NEXT: v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v32, v26
-; VI-NEXT: v_add_u16_e32 v32, 3, v25
-; VI-NEXT: v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v32, v25
-; VI-NEXT: v_add_u16_e32 v32, 3, v24
-; VI-NEXT: v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v32, v24
-; VI-NEXT: v_add_u16_e32 v32, 3, v23
-; VI-NEXT: v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v32, v23
-; VI-NEXT: v_add_u16_e32 v32, 3, v22
-; VI-NEXT: v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v32, v22
-; VI-NEXT: v_add_u16_e32 v32, 3, v21
-; VI-NEXT: v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v32, v21
-; VI-NEXT: v_add_u16_e32 v32, 3, v20
-; VI-NEXT: v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v32, v20
-; VI-NEXT: v_add_u16_e32 v32, 3, v19
-; VI-NEXT: v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v32, v19
-; VI-NEXT: v_add_u16_e32 v32, 3, v18
-; VI-NEXT: v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v32, v18
-; VI-NEXT: v_add_u16_e32 v32, 3, v17
-; VI-NEXT: v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v32, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB13_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v32i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB13_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v32i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB13_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB13_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <32 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <32 x i32>
- br label %end
-
-end:
- %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i32> %phi
-}
-
-define <16 x i64> @bitcast_v32f32_to_v16i64(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB14_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB14_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB14_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB14_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB14_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB14_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <32 x float> @bitcast_v16i64_to_v32f32(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB15_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB15_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB15_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB15_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB15_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB15_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <16 x double> @bitcast_v32f32_to_v16f64(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB16_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB16_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB16_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB16_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB16_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB16_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <32 x float> @bitcast_v16f64_to_v32f32(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GCN-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GCN-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GCN-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GCN-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GCN-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GCN-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GCN-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB18_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_f32_e32 v58, 1.0, v58
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v59, 1.0, v59
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v58, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v58
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB18_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v52
-; GCN-NEXT: v_or_b32_e32 v1, v1, v52
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_or_b32_e32 v2, v2, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v50, v31
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v31
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v50
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v49
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v61
-; GCN-NEXT: v_or_b32_e32 v49, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v48
-; GCN-NEXT: v_or_b32_e32 v2, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v60
-; GCN-NEXT: v_or_b32_e32 v61, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v39
-; GCN-NEXT: v_or_b32_e32 v62, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v57
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v38
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v56
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v37
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v47
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v36
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v46
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v35
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v45
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v34
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v44
-; GCN-NEXT: v_or_b32_e32 v14, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v33
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v43
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v21
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v42
-; GCN-NEXT: v_or_b32_e32 v18, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v41
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v25
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v40
-; GCN-NEXT: v_or_b32_e32 v22, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v27
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v55
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v25, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v26, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v27, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v53
-; GCN-NEXT: v_or_b32_e32 v28, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v29, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v30, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v1, v29
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v49
-; GCN-NEXT: v_or_b32_e32 v5, v1, v32
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v30
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v62
-; GCN-NEXT: v_or_b32_e32 v6, v6, v34
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v36
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v37
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v38
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v48
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v9, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v11, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v12, v12, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v13, v13, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v15, v15, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v17, v17, v42
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v43
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v19, v44
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v45
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v21, v21, v46
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v47
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v23, v23, v56
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v57
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v25, v25, v58
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v59
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v27, v27, v60
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB18_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v39, 24, v32
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_mov_b32_e32 v55, v39
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: .LBB18_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB18_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v32, 1.0, v32
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: .LBB18_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v48
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v57
-; VI-NEXT: v_or_b32_sdwa v2, v2, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v41
-; VI-NEXT: v_or_b32_sdwa v48, v53, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v38
-; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v47
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v45
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v46, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v37
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v49
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB18_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_waitcnt vmcnt(29)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(31)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: .LBB18_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB18_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(29)
-; GFX9-NEXT: v_add_f32_e32 v32, 1.0, v32
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: .LBB18_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v52
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v56
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v47
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v46
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v41
-; GFX9-NEXT: v_or_b32_sdwa v33, v44, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_or_b32_sdwa v1, v53, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB18_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: .LBB18_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB18_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v20, 1.0, v20 :: v_dual_add_f32 v19, 1.0, v19
-; GFX11-NEXT: v_dual_add_f32 v18, 1.0, v18 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v24, 1.0, v24 :: v_dual_add_f32 v31, 1.0, v31
-; GFX11-NEXT: v_dual_add_f32 v32, 1.0, v32 :: v_dual_add_f32 v23, 1.0, v23
-; GFX11-NEXT: v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-NEXT: v_dual_add_f32 v26, 1.0, v26 :: v_dual_add_f32 v29, 1.0, v29
-; GFX11-NEXT: v_dual_add_f32 v30, 1.0, v30 :: v_dual_add_f32 v25, 1.0, v25
-; GFX11-NEXT: v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v13, 1.0, v13
-; GFX11-NEXT: v_dual_add_f32 v28, 1.0, v28 :: v_dual_add_f32 v27, 1.0, v27
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v11, 1.0, v11
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v9, 1.0, v9
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v7, 1.0, v7
-; GFX11-NEXT: v_dual_add_f32 v22, 1.0, v22 :: v_dual_add_f32 v21, 1.0, v21
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: .LBB18_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v63
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v65
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v61
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v73
-; GFX11-NEXT: v_or_b32_e32 v65, v67, v65
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v66
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v62
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v66
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v60
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v65
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v67
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v56
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v65
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v47
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v45
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v64
-; GFX11-NEXT: v_or_b32_e32 v55, v65, v66
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v44
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v43
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v54, v65, v54
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v40
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v55
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v65
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v182
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v181
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v55
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v65
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v54
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v53, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v176
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v167
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v53, v53, v54
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v52
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v65
-; GFX11-NEXT: v_or_b32_e32 v54, v66, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v53
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v54
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v164
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v160
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v150
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v146
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v49
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v145
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v144
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v135
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v132
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v131
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v128
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v119
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v117
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v116
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v115
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v103
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v102
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v101
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v99
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v86
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v84
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v83
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v81
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v70
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v69
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v68
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v10
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:148
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:160
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:192
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:212
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:244
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:288
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:340
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:352
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:380
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:376
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v1
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:372
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:364
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v0, v39
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v47
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v54
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v46
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v37
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v48
-; GCN-NEXT: v_or_b32_e32 v8, v8, v45
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v53
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v42
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v41
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v63
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v50
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v49
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v20, v60
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v58
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v22, v62
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v59
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v61
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v52
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v35
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v33
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v44
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v57
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v38, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v48, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v55, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v54, v49
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v18, v18, v50
-; GCN-NEXT: v_or_b32_e32 v19, v19, v51
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_or_b32_e32 v21, v21, v53
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: v_or_b32_e32 v23, v23, v33
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_or_b32_e32 v25, v25, v35
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v37
-; GCN-NEXT: v_or_b32_e32 v28, v28, v38
-; GCN-NEXT: v_or_b32_e32 v29, v29, v39
-; GCN-NEXT: v_or_b32_e32 v30, v30, v48
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; kill: killed $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB19_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v39, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v56, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v47, v3
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v54, v4
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v46, v5
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v37, v6
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v43, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v45, v8
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v42, v10
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v41, v11
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v40, v12
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v63, v13
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v50, v14
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v0, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v0, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v0, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v51, v18
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v49, v19
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v60, v20
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v58, v21
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v25, v62, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v29, v59, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v37, v32, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v50, v61, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v41, v34, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v45, v52, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v56, v35, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v58, v33, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v59, v44, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v57
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v57, v36, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v26, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v48, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v53, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_mov_b32_e32 v0, v55
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v55, v53
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v40, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v40
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v42, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v44, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v46, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v46
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_or_b32_e32 v46, v0, v46
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v47
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v0, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v60, v0
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v61, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v62, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v63, v3
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s7, v37
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s7, v50
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s7, v41
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s7, v45
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s7, v56
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 0x300, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff, v56
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_or_b32_e32 v4, v36, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v39, v6
-; GCN-NEXT: v_or_b32_e32 v7, v49, v7
-; GCN-NEXT: v_or_b32_e32 v8, v51, v8
-; GCN-NEXT: v_or_b32_e32 v9, v52, v9
-; GCN-NEXT: v_or_b32_e32 v10, v54, v10
-; GCN-NEXT: v_or_b32_e32 v11, v22, v11
-; GCN-NEXT: v_or_b32_e32 v12, v23, v12
-; GCN-NEXT: v_or_b32_e32 v13, v24, v13
-; GCN-NEXT: v_or_b32_e32 v14, v26, v14
-; GCN-NEXT: v_or_b32_e32 v15, v27, v15
-; GCN-NEXT: v_or_b32_e32 v16, v28, v16
-; GCN-NEXT: v_or_b32_e32 v17, v30, v17
-; GCN-NEXT: v_or_b32_e32 v18, v31, v18
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: v_or_b32_e32 v20, v33, v20
-; GCN-NEXT: v_or_b32_e32 v21, v34, v21
-; GCN-NEXT: v_or_b32_e32 v22, v35, v25
-; GCN-NEXT: v_or_b32_e32 v23, v48, v29
-; GCN-NEXT: v_or_b32_e32 v24, v53, v37
-; GCN-NEXT: v_or_b32_e32 v25, v55, v50
-; GCN-NEXT: v_or_b32_e32 v26, v40, v41
-; GCN-NEXT: v_or_b32_e32 v27, v42, v45
-; GCN-NEXT: v_or_b32_e32 v28, v43, v56
-; GCN-NEXT: v_or_b32_e32 v29, v44, v58
-; GCN-NEXT: v_or_b32_e32 v30, v46, v59
-; GCN-NEXT: v_or_b32_e32 v31, v47, v57
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x3000000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x3000000, v31
-; GCN-NEXT: .LBB19_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: .LBB19_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v31, 0x300
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v61
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v63
-; VI-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v59
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v62
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v58
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v60
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v54
-; VI-NEXT: v_add_u16_e32 v15, 3, v40
-; VI-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v14, v15
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v15, v16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v16, v16, v17
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v18
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v18, 0x300, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v18, v19
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v19, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 0x300, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v20, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v21, 0x300, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v21, v22
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v22, v23
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v23, v24
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v24, v25
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v25, v26
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v26, v27
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v27, v28
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v28, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v32, 0x300, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v33, 3, v33
-; VI-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: .LBB19_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB19_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: .LBB19_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB19_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v54
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v29, v29, v30
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v30, v30, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v31, v31, v32
-; GFX9-NEXT: .LBB19_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:592
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:588
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:584
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:468
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:456
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v136, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v137, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v138, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
-; GFX11-NEXT: v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
-; GFX11-NEXT: v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
-; GFX11-NEXT: v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
-; GFX11-NEXT: v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
-; GFX11-NEXT: v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
-; GFX11-NEXT: v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v114, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v115, off, s32
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v181, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v182, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v136, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v137, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v138, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v132, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v150, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v160, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v161, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v167, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v176, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v41, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v115
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v116
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v117
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v118
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v119
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v131
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v144
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v145
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v146
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v147
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v148
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v162
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v163
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v164
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v165
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v166
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v179, 8, v179
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v180, 8, v180
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v181
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v183, 8, v183
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v162, 8, v136
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v163, 8, v137
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v164, 8, v138
-; GFX11-NEXT: v_lshlrev_b16 v165, 8, v103
-; GFX11-NEXT: v_lshlrev_b16 v166, 8, v102
-; GFX11-NEXT: v_lshlrev_b16 v144, 8, v101
-; GFX11-NEXT: v_lshlrev_b16 v145, 8, v100
-; GFX11-NEXT: v_lshlrev_b16 v146, 8, v99
-; GFX11-NEXT: v_lshlrev_b16 v147, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v148, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v28
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v26
-; GFX11-NEXT: v_lshlrev_b16 v129, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v130, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v131, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v14
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v0
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB19_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v51
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v124
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v125
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v126
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v127
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v111
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v121
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v120
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v122
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v123
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v107
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v108
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v109
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v110
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v106
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v7, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v6, v12
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v92
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v78
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v77
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v76
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v75
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v59
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v93
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v94
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v95
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v104
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v105
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v79
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v88
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v89
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v90
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v91
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v58
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v44
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v43
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v41
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v40
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v176
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v167
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v61
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v62
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v63
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v72
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v73
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v45
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v46
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v47
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v56
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v57
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v161
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v160
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v150
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v132
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v113
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v179
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v180
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v181
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v182
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v183
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v162
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v163
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v164
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v165
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v166
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v84
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v82
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v144
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v145
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v146
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v147
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v148
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v119
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v128
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v129
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v130
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v131
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v66
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v64
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v114
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v115
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v116
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v117
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v118
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v99
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v100
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v101
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v102
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v103
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: .LBB19_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB19_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v51, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v5, v50, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v124, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v125, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v126, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v127, v3
-; GFX11-NEXT: v_add_nc_u16 v6, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v7, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_add_nc_u16 v8, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v36, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v10, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v2, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v11, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_or_b32_e32 v3, v111, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v120, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v121, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v6, v122, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v107, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v108, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v109, v10
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v38, 3
-; GFX11-NEXT: v_or_b32_e32 v11, v110, v11
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v106, v2
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_add_nc_u16 v7, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v32, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v9, v92, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v78, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v77, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v76, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v75, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v74, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v60, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v59, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v93, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v94, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v95, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v104, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v105, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v79, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v88, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v89, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v90, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v91, v16
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_add_nc_u16 v12, v58, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v44, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v43, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v41, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v40, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v178, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v177, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v176, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v167, 3
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v61, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v62, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v63, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v72, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v73, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v45, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v46, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v47, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v56, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v57, v21
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_add_nc_u16 v17, v161, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v160, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v150, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v149, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v133, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v132, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v113, 3
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v179, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v180, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v181, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v182, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v183, v21
-; GFX11-NEXT: v_or_b32_e32 v22, v162, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v163, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v164, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v165, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v166, v26
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_add_nc_u16 v22, v112, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v97, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v87, 3
-; GFX11-NEXT: v_add_nc_u16 v27, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v84, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v83, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v82, 3
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v144, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v145, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v146, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v147, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v148, v26
-; GFX11-NEXT: v_or_b32_e32 v27, v119, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v128, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v129, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v130, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v131, v31
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_add_nc_u16 v27, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v33, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v34, v66, 3
-; GFX11-NEXT: v_add_nc_u16 v35, v65, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v64, 3
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v114, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v115, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v116, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v117, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v118, v31
-; GFX11-NEXT: v_or_b32_e32 v32, v99, v32
-; GFX11-NEXT: v_or_b32_e32 v33, v100, v33
-; GFX11-NEXT: v_or_b32_e32 v34, v101, v34
-; GFX11-NEXT: v_or_b32_e32 v35, v102, v35
-; GFX11-NEXT: v_or_b32_e32 v36, v103, v36
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v34
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v36
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: .LBB19_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v138, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v137, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v136, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:516
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:580
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:584
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:588
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:592
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v62
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: .LBB20_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_add_f32_e32 v32, 1.0, v62
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: .LBB20_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v61
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v59
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v56, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v47
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v46, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v45, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v44, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 20, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v33
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v17, 16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v18, v18, v19, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v20, v20, v21, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_alignbit_b32 v24, v24, v25, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_alignbit_b32 v26, v26, v27, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_alignbit_b32 v28, v28, v29, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_alignbit_b32 v30, v30, v31, 16
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_alignbit_b32 v32, v32, v33, 16
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_alignbit_b32 v34, v34, v35, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_alignbit_b32 v36, v36, v37, 16
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v38, v38, v39, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_alignbit_b32 v48, v48, v49, 16
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_alignbit_b32 v50, v50, v51, 16
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_alignbit_b32 v52, v52, v53, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_alignbit_b32 v54, v54, v55, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_alignbit_b32 v40, v40, v41, 16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v63
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_alignbit_b32 v42, v42, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v56, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB20_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: .LBB20_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB20_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: .LBB20_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB20_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: .LBB20_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <32 x float> @bitcast_v64bf16_to_v32f32(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v33
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v42
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v51
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v63
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v60
-; GCN-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v61, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v58
-; GCN-NEXT: v_alignbit_b32 v2, v2, v59, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v56
-; GCN-NEXT: v_alignbit_b32 v3, v3, v57, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v46
-; GCN-NEXT: v_alignbit_b32 v4, v4, v47, 16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v45, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v11, v11, v12, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v13, v13, v14, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v34
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v33, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v17, v18, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v43
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v19, v19, v42, 16
-; GCN-NEXT: v_alignbit_b32 v20, v20, v44, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; GCN-NEXT: v_alignbit_b32 v22, v22, v48, 16
-; GCN-NEXT: v_alignbit_b32 v23, v23, v38, 16
-; GCN-NEXT: v_alignbit_b32 v24, v24, v50, 16
-; GCN-NEXT: v_alignbit_b32 v25, v25, v51, 16
-; GCN-NEXT: v_alignbit_b32 v26, v26, v53, 16
-; GCN-NEXT: v_alignbit_b32 v27, v27, v55, 16
-; GCN-NEXT: v_alignbit_b32 v28, v28, v40, 16
-; GCN-NEXT: v_alignbit_b32 v29, v29, v63, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; kill: killed $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: .LBB21_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v4, v5, v4, 16
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_alignbit_b32 v6, v7, v6, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v7, v8, v7, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v8, v9, v8, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_alignbit_b32 v10, v11, v10, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v11, v12, v11, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v12, v13, v12, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_alignbit_b32 v14, v15, v14, 16
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v34
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v15, v16, v15, 16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v33
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_alignbit_b32 v16, v17, v16, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v42
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v44
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v54
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v41
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v43
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v40, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v44, 0x40c00000, v31
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v32
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v33
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v18, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v19, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v20, v39, v54, 16
-; GCN-NEXT: v_alignbit_b32 v21, v48, v40, 16
-; GCN-NEXT: v_alignbit_b32 v22, v49, v22, 16
-; GCN-NEXT: v_alignbit_b32 v23, v50, v23, 16
-; GCN-NEXT: v_alignbit_b32 v24, v51, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v32, v25, 16
-; GCN-NEXT: v_alignbit_b32 v26, v33, v26, 16
-; GCN-NEXT: v_alignbit_b32 v27, v34, v27, 16
-; GCN-NEXT: v_alignbit_b32 v28, v35, v28, 16
-; GCN-NEXT: v_alignbit_b32 v29, v36, v29, 16
-; GCN-NEXT: v_alignbit_b32 v30, v37, v30, 16
-; GCN-NEXT: v_alignbit_b32 v31, v38, v31, 16
-; GCN-NEXT: .LBB21_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v15
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v14
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v13
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v12
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v11
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v10
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v9
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v8
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v7
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v6
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v5
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v4
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v3
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v2
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v1
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v0
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v31
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v30
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v29
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v29, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v28
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v28, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v27
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v27, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v26
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v26, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v25
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v25, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v24
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v24, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v23
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v23, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v22
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v22, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v21
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v20
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v20, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v19
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v19, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v18
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v17
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v17, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB21_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v32, s7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v31, v31, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v30, v30, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v29, v29, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v28, v28, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v27, v27, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v26, v26, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v25, v25, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v24, v24, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v23, v23, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v22, v22, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v21, v21, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v20, v20, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v19, v19, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v18, v18, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v17, v17, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v16, v16, v32, s7
-; GFX9-NEXT: .LBB21_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB21_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v32, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v32
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v39, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_add3_u32 v34, v34, v32, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v39, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v36, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v36, v36, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v32, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_perm_b32 v14, v14, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v33, v33, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v37, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v13, v13, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v34, v34, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v33, v33, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_bfe_u32 v35, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v35, v35, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v33, v33, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v35, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v33, v33, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GFX11-NEXT: v_perm_b32 v7, v7, v32, 0x7060302
-; GFX11-NEXT: v_perm_b32 v6, v6, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_bfe_u32 v34, v4, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v5, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v3
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v4
-; GFX11-NEXT: v_perm_b32 v5, v5, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v4, v4, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v35, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v2, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_perm_b32 v3, v3, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_perm_b32 v2, v2, v33, 0x7060302
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v34, v0, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v0
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_perm_b32 v1, v1, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v0, v0, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX11-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_bfe_u32 v35, v30, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v29
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v30, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_perm_b32 v31, v31, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v30, v30, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_add3_u32 v33, v33, v29, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v29
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-NEXT: v_bfe_u32 v34, v28, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v27
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v28, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v28
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v26
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_perm_b32 v29, v29, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v28, v28, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX11-NEXT: v_add3_u32 v33, v33, v27, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v27
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-NEXT: v_bfe_u32 v35, v26, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v25
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v26, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v26
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_perm_b32 v27, v27, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v26, v26, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-NEXT: v_add3_u32 v33, v33, v25, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v25
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-NEXT: v_bfe_u32 v34, v24, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v23
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v24
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_perm_b32 v25, v25, v32, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v24, v24, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v33, v23, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-NEXT: v_bfe_u32 v35, v22, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v22, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v22
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v23, v23, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v36, v34, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v21, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v21
-; GFX11-NEXT: v_perm_b32 v22, v22, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GFX11-NEXT: v_add3_u32 v35, v36, v34, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v21, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v33, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v32, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v32, v38, v33, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v21, v21, v34, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v34, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v35, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_add3_u32 v34, v34, v20, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v34, v35, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v36
-; GFX11-NEXT: v_add3_u32 v35, v37, v33, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v33
-; GFX11-NEXT: v_bfe_u32 v37, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_perm_b32 v20, v20, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v37, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v37, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_bfe_u32 v48, v18, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v36, v48, v18, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX11-NEXT: v_perm_b32 v19, v19, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v36, v37, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-NEXT: v_add3_u32 v37, v38, v35, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v35
-; GFX11-NEXT: v_bfe_u32 v39, v17, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_bfe_u32 v48, v36, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v36
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v16
-; GFX11-NEXT: v_perm_b32 v18, v18, v34, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v39, v17, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_bfe_u32 v37, v16, 16, 1
-; GFX11-NEXT: v_add3_u32 v48, v48, v36, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v38, v39, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_add3_u32 v37, v37, v16, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v17, v17, v35, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v48, v49, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v37, v50, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v16, v16, v36, 0x7060302
-; GFX11-NEXT: .LBB21_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <64 x half> @bitcast_v32f32_to_v64f16(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v62
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v62
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v63
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v36
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v38
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v48
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v49
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v50
-; GCN-NEXT: v_mov_b32_e32 v50, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB22_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v31, v33
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v63
-; GCN-NEXT: v_add_f32_e32 v32, 1.0, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v61
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: .LBB22_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v59
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v58
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v2, v1
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v57
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v2, v1
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 12, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v47
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v2, v1
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v46
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 20, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v45
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v44
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v43
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v42
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v41
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v54
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v53
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v49
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v48
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v36
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v49, v48
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v52, v53, v52
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v55, v54
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v40, v41, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v61, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v60, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v46, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB22_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: .LBB22_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB22_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: .LBB22_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB22_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: .LBB22_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <32 x float> @bitcast_v64f16_to_v32f32(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v43
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v50
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v1
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v61
-; GCN-NEXT: v_or_b32_e32 v0, v62, v0
-; GCN-NEXT: v_or_b32_e32 v1, v60, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: v_or_b32_e32 v2, v58, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v57
-; GCN-NEXT: v_or_b32_e32 v3, v56, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v47
-; GCN-NEXT: v_or_b32_e32 v4, v46, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v32
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v37
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_or_b32_e32 v19, v43, v19
-; GCN-NEXT: v_or_b32_e32 v20, v41, v20
-; GCN-NEXT: v_or_b32_e32 v21, v55, v21
-; GCN-NEXT: v_or_b32_e32 v22, v49, v22
-; GCN-NEXT: v_or_b32_e32 v23, v50, v23
-; GCN-NEXT: v_or_b32_e32 v24, v39, v24
-; GCN-NEXT: v_or_b32_e32 v25, v36, v25
-; GCN-NEXT: v_or_b32_e32 v26, v48, v26
-; GCN-NEXT: v_or_b32_e32 v27, v52, v27
-; GCN-NEXT: v_or_b32_e32 v28, v53, v28
-; GCN-NEXT: v_or_b32_e32 v29, v54, v29
-; GCN-NEXT: v_or_b32_e32 v30, v40, v30
-; GCN-NEXT: v_or_b32_e32 v31, v42, v31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: .LBB23_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v41
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v55
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v50
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v39
-; GCN-NEXT: v_mov_b32_e32 v39, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v48
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v33
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_or_b32_e32 v19, v21, v20
-; GCN-NEXT: v_or_b32_e32 v20, v55, v39
-; GCN-NEXT: v_or_b32_e32 v21, v41, v48
-; GCN-NEXT: v_or_b32_e32 v22, v22, v49
-; GCN-NEXT: v_or_b32_e32 v23, v23, v50
-; GCN-NEXT: v_or_b32_e32 v24, v24, v51
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v38
-; GCN-NEXT: v_or_b32_e32 v28, v28, v33
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_or_b32_e32 v30, v30, v35
-; GCN-NEXT: v_or_b32_e32 v31, v31, v37
-; GCN-NEXT: .LBB23_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB23_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB23_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB23_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB23_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB23_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB23_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <64 x i16> @bitcast_v32f32_to_v64i16(<32 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f32_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; kill: killed $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; kill: killed $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v50, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v52, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v55, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v41, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB24_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v32, 1.0, v32
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v50, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v52, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v55, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v41, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB24_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v59
-; GCN-NEXT: v_or_b32_e32 v1, v1, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v33
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v57
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v46
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v62
-; GCN-NEXT: v_or_b32_e32 v46, v1, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v43
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v41
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v60
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v55
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v58
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v52
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v56
-; GCN-NEXT: v_or_b32_e32 v14, v14, v33
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v15, v15, v50
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v47
-; GCN-NEXT: v_or_b32_e32 v16, v16, v52
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v17, v17, v48
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v45
-; GCN-NEXT: v_or_b32_e32 v18, v18, v55
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_or_b32_e32 v19, v19, v39
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v20, v20, v34
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v38
-; GCN-NEXT: v_or_b32_e32 v21, v21, v34
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v42
-; GCN-NEXT: v_or_b32_e32 v22, v22, v34
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v37
-; GCN-NEXT: v_or_b32_e32 v23, v23, v34
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v40
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v36
-; GCN-NEXT: v_or_b32_e32 v25, v25, v34
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v54
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v35
-; GCN-NEXT: v_or_b32_e32 v27, v27, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v53
-; GCN-NEXT: v_or_b32_e32 v28, v28, v34
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_or_b32_e32 v30, v30, v51
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_or_b32_e32 v32, v32, v43
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v44, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f32_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f32_e32 v31, 1.0, v31
-; VI-NEXT: v_add_f32_e32 v30, 1.0, v30
-; VI-NEXT: v_add_f32_e32 v29, 1.0, v29
-; VI-NEXT: v_add_f32_e32 v28, 1.0, v28
-; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
-; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
-; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
-; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
-; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
-; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
-; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
-; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: .LBB24_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f32_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f32_e32 v31, 1.0, v31
-; GFX9-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GFX9-NEXT: v_add_f32_e32 v29, 1.0, v29
-; GFX9-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
-; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
-; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
-; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
-; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: .LBB24_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f32_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-NEXT: v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: .LBB24_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <32 x float> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x float> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v32f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v20
-; GCN-NEXT: v_mov_b32_e32 v38, v18
-; GCN-NEXT: v_mov_b32_e32 v39, v16
-; GCN-NEXT: v_mov_b32_e32 v48, v14
-; GCN-NEXT: v_mov_b32_e32 v49, v12
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v51, v8
-; GCN-NEXT: v_mov_b32_e32 v52, v6
-; GCN-NEXT: v_mov_b32_e32 v53, v4
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(12) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v24
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v55
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v36
-; GCN-NEXT: v_or_b32_e32 v1, v1, v58
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v53
-; GCN-NEXT: v_or_b32_e32 v2, v2, v57
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v52
-; GCN-NEXT: v_or_b32_e32 v3, v3, v35
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v4, v4, v60
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v50
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v49
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v48
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v39
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v38
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v43
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v46
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v40
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v62
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v47
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v44
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v18, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v21, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v29, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v30, v32
-; GCN-NEXT: v_or_b32_e32 v31, v31, v59
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: .LBB25_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v36, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v58, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v57, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v52
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v35, v3
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v60, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v32, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v32, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v15
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v32, v16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v32, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v32, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v32, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v32, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v32, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v32, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v32, v30
-; GCN-NEXT: v_or_b32_e32 v31, v59, v31
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x30000, v31
-; GCN-NEXT: .LBB25_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(11)
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v32f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v33, 3
-; VI-NEXT: v_add_u16_e32 v32, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v32, v15
-; VI-NEXT: v_add_u16_e32 v32, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v32, v14
-; VI-NEXT: v_add_u16_e32 v32, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v32, v13
-; VI-NEXT: v_add_u16_e32 v32, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v32, v12
-; VI-NEXT: v_add_u16_e32 v32, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v32, v11
-; VI-NEXT: v_add_u16_e32 v32, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v32, v10
-; VI-NEXT: v_add_u16_e32 v32, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v32, v9
-; VI-NEXT: v_add_u16_e32 v32, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v32, v8
-; VI-NEXT: v_add_u16_e32 v32, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v32, v7
-; VI-NEXT: v_add_u16_e32 v32, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v32, v6
-; VI-NEXT: v_add_u16_e32 v32, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v32, v5
-; VI-NEXT: v_add_u16_e32 v32, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v32, v4
-; VI-NEXT: v_add_u16_e32 v32, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v32, v3
-; VI-NEXT: v_add_u16_e32 v32, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v32, v2
-; VI-NEXT: v_add_u16_e32 v32, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v32, v1
-; VI-NEXT: v_add_u16_e32 v32, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v32, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v31
-; VI-NEXT: v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: v_add_u16_e32 v32, 3, v30
-; VI-NEXT: v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v32, v30
-; VI-NEXT: v_add_u16_e32 v32, 3, v29
-; VI-NEXT: v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v32, v29
-; VI-NEXT: v_add_u16_e32 v32, 3, v28
-; VI-NEXT: v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v32, v28
-; VI-NEXT: v_add_u16_e32 v32, 3, v27
-; VI-NEXT: v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_add_u16_e32 v32, 3, v26
-; VI-NEXT: v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v32, v26
-; VI-NEXT: v_add_u16_e32 v32, 3, v25
-; VI-NEXT: v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v32, v25
-; VI-NEXT: v_add_u16_e32 v32, 3, v24
-; VI-NEXT: v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v32, v24
-; VI-NEXT: v_add_u16_e32 v32, 3, v23
-; VI-NEXT: v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v32, v23
-; VI-NEXT: v_add_u16_e32 v32, 3, v22
-; VI-NEXT: v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v32, v22
-; VI-NEXT: v_add_u16_e32 v32, 3, v21
-; VI-NEXT: v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v32, v21
-; VI-NEXT: v_add_u16_e32 v32, 3, v20
-; VI-NEXT: v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v32, v20
-; VI-NEXT: v_add_u16_e32 v32, 3, v19
-; VI-NEXT: v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v32, v19
-; VI-NEXT: v_add_u16_e32 v32, 3, v18
-; VI-NEXT: v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v32, v18
-; VI-NEXT: v_add_u16_e32 v32, 3, v17
-; VI-NEXT: v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v32, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB25_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v32f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB25_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v32f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB25_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <32 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <32 x float>
- br label %end
-
-end:
- %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x float> %phi
-}
-
-define <16 x double> @bitcast_v16i64_to_v16f64(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; GCN-NEXT: .LBB26_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB26_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: .LBB26_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB26_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: .LBB26_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB26_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: .LBB26_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <16 x i64> @bitcast_v16f64_to_v16i64(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GCN-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GCN-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GCN-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GCN-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GCN-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GCN-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GCN-NEXT: .LBB27_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: .LBB27_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: .LBB27_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: .LBB27_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v57
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v57
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB28_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 3, v59
-; GCN-NEXT: v_addc_u32_e32 v57, vcc, 0, v57, vcc
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v57, v59, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v34, v18, v17, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v14, v13, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v12, v11, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v10, v9, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v50, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v51, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v52, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v57
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v57
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 8, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v28
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v26
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v20
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v18
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v16
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v12
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v10
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v8
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v6
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v4
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; GCN-NEXT: .LBB28_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v52
-; GCN-NEXT: v_or_b32_e32 v1, v1, v52
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_or_b32_e32 v2, v2, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v50, v31
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v31
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v50
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v49
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v61
-; GCN-NEXT: v_or_b32_e32 v49, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v48
-; GCN-NEXT: v_or_b32_e32 v2, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v60
-; GCN-NEXT: v_or_b32_e32 v61, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v39
-; GCN-NEXT: v_or_b32_e32 v62, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v58
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v38
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v56
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v37
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v47
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v36
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v46
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v35
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v45
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v34
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v44
-; GCN-NEXT: v_or_b32_e32 v14, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v33
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v43
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v21
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v42
-; GCN-NEXT: v_or_b32_e32 v18, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v41
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v25
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v40
-; GCN-NEXT: v_or_b32_e32 v22, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v27
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v55
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v25, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v26, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v27, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v53
-; GCN-NEXT: v_or_b32_e32 v28, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v29, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v30, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v1, v29
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v49
-; GCN-NEXT: v_or_b32_e32 v5, v1, v32
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v30
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v62
-; GCN-NEXT: v_or_b32_e32 v6, v6, v34
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v36
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v37
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v38
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v48
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v9, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v11, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v12, v12, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v13, v13, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v15, v15, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v17, v17, v42
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v43
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v19, v44
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v45
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v21, v21, v46
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v47
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v23, v23, v56
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v57
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v25, v25, v58
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v59
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v27, v27, v60
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB28_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v39, 24, v32
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_mov_b32_e32 v55, v39
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: .LBB28_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB28_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
-; VI-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
-; VI-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
-; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
-; VI-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
-; VI-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v29
-; VI-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; VI-NEXT: v_add_u32_e32 v31, vcc, 3, v31
-; VI-NEXT: v_addc_u32_e32 v32, vcc, 0, v32, vcc
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v27
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v10
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v34, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v1
-; VI-NEXT: .LBB28_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v48
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v57
-; VI-NEXT: v_or_b32_sdwa v2, v2, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v41
-; VI-NEXT: v_or_b32_sdwa v48, v53, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v38
-; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v47
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v45
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v46, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v37
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v49
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; kill: killed $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB28_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_waitcnt vmcnt(29)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(31)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: .LBB28_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB28_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v1, vcc, 3, v1
-; GFX9-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, 3, v3
-; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, 3, v5
-; GFX9-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v7, vcc, 3, v7
-; GFX9-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v8, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v9, vcc, 3, v9
-; GFX9-NEXT: v_addc_co_u32_e32 v10, vcc, 0, v10, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v11, vcc, 3, v11
-; GFX9-NEXT: v_addc_co_u32_e32 v12, vcc, 0, v12, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v13, vcc, 3, v13
-; GFX9-NEXT: v_addc_co_u32_e32 v14, vcc, 0, v14, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v15, vcc, 3, v15
-; GFX9-NEXT: v_addc_co_u32_e32 v16, vcc, 0, v16, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v17, vcc, 3, v17
-; GFX9-NEXT: v_addc_co_u32_e32 v18, vcc, 0, v18, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v19, vcc, 3, v19
-; GFX9-NEXT: v_addc_co_u32_e32 v20, vcc, 0, v20, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v21, vcc, 3, v21
-; GFX9-NEXT: v_addc_co_u32_e32 v22, vcc, 0, v22, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v23, vcc, 3, v23
-; GFX9-NEXT: v_addc_co_u32_e32 v24, vcc, 0, v24, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v25, vcc, 3, v25
-; GFX9-NEXT: v_addc_co_u32_e32 v26, vcc, 0, v26, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v27, vcc, 3, v27
-; GFX9-NEXT: v_addc_co_u32_e32 v28, vcc, 0, v28, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v29, vcc, 3, v29
-; GFX9-NEXT: v_addc_co_u32_e32 v30, vcc, 0, v30, vcc
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_co_u32_e32 v31, vcc, 3, v31
-; GFX9-NEXT: v_addc_co_u32_e32 v32, vcc, 0, v32, vcc
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v10
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v1
-; GFX9-NEXT: .LBB28_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v52
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v56
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v47
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v46
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v41
-; GFX9-NEXT: v_or_b32_sdwa v33, v44, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_or_b32_sdwa v1, v53, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: .LBB28_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v1, vcc_lo, v1, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, v3, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, v5, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v7, vcc_lo, v7, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v9, vcc_lo, v9, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v11, vcc_lo, v11, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v12, null, 0, v12, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v13, vcc_lo, v13, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v14, null, 0, v14, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v15, vcc_lo, v15, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v16, null, 0, v16, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v17, vcc_lo, v17, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v18, null, 0, v18, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v19, vcc_lo, v19, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v20, null, 0, v20, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v21, vcc_lo, v21, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v22, null, 0, v22, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v25, vcc_lo, v25, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v26, null, 0, v26, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v27, vcc_lo, v27, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v28, null, 0, v28, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v29, vcc_lo, v29, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v30, null, 0, v30, vcc_lo
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_u32 v31, vcc_lo, v31, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v32, null, 0, v32, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v23, vcc_lo, v23, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v24, null, 0, v24, vcc_lo
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: .LBB28_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v63
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v65
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v61
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v73
-; GFX11-NEXT: v_or_b32_e32 v65, v67, v65
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v66
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v62
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v66
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v60
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v65
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v67
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v56
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v65
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v47
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v45
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v64
-; GFX11-NEXT: v_or_b32_e32 v55, v65, v66
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v44
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v43
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v54, v65, v54
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v40
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v55
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v65
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v182
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v181
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v55
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v65
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v54
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v53, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v176
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v167
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v53, v53, v54
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v52
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v65
-; GFX11-NEXT: v_or_b32_e32 v54, v66, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v53
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v54
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v164
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v160
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v150
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v146
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v49
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v145
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v144
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v135
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v132
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v131
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v128
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v119
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v117
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v116
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v115
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v103
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v102
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v101
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v99
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v86
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v84
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v83
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v81
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v70
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v69
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v68
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v10
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:148
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:160
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:192
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:212
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:244
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:288
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:340
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:352
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:380
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:376
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v1
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:372
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:364
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v0, v39
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v47
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v54
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v46
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v37
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v48
-; GCN-NEXT: v_or_b32_e32 v8, v8, v45
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v53
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v42
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v41
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v63
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v50
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v49
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v20, v60
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v58
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v22, v62
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v59
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v61
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v52
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v35
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v33
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v44
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v57
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v38, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v48, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v55, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v54, v49
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v18, v18, v50
-; GCN-NEXT: v_or_b32_e32 v19, v19, v51
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_or_b32_e32 v21, v21, v53
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: v_or_b32_e32 v23, v23, v33
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_or_b32_e32 v25, v25, v35
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v37
-; GCN-NEXT: v_or_b32_e32 v28, v28, v38
-; GCN-NEXT: v_or_b32_e32 v29, v29, v39
-; GCN-NEXT: v_or_b32_e32 v30, v30, v48
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; kill: killed $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB29_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v39, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v56, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v47, v3
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v54, v4
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v46, v5
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v37, v6
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v43, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v45, v8
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v42, v10
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v41, v11
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v40, v12
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v63, v13
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v50, v14
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v0, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v0, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v0, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v51, v18
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v49, v19
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v60, v20
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v58, v21
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v25, v62, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v29, v59, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v37, v32, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v50, v61, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v41, v34, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v45, v52, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v56, v35, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v58, v33, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v59, v44, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v57
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v57, v36, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v26, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v48, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v53, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_mov_b32_e32 v0, v55
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v55, v53
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v40, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v40
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v42, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v44, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v46, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v46
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_or_b32_e32 v46, v0, v46
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v47
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v0, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v60, v0
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v61, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v62, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v63, v3
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s7, v37
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s7, v50
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s7, v41
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s7, v45
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s7, v56
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 0x300, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff, v56
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_or_b32_e32 v4, v36, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v39, v6
-; GCN-NEXT: v_or_b32_e32 v7, v49, v7
-; GCN-NEXT: v_or_b32_e32 v8, v51, v8
-; GCN-NEXT: v_or_b32_e32 v9, v52, v9
-; GCN-NEXT: v_or_b32_e32 v10, v54, v10
-; GCN-NEXT: v_or_b32_e32 v11, v22, v11
-; GCN-NEXT: v_or_b32_e32 v12, v23, v12
-; GCN-NEXT: v_or_b32_e32 v13, v24, v13
-; GCN-NEXT: v_or_b32_e32 v14, v26, v14
-; GCN-NEXT: v_or_b32_e32 v15, v27, v15
-; GCN-NEXT: v_or_b32_e32 v16, v28, v16
-; GCN-NEXT: v_or_b32_e32 v17, v30, v17
-; GCN-NEXT: v_or_b32_e32 v18, v31, v18
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: v_or_b32_e32 v20, v33, v20
-; GCN-NEXT: v_or_b32_e32 v21, v34, v21
-; GCN-NEXT: v_or_b32_e32 v22, v35, v25
-; GCN-NEXT: v_or_b32_e32 v23, v48, v29
-; GCN-NEXT: v_or_b32_e32 v24, v53, v37
-; GCN-NEXT: v_or_b32_e32 v25, v55, v50
-; GCN-NEXT: v_or_b32_e32 v26, v40, v41
-; GCN-NEXT: v_or_b32_e32 v27, v42, v45
-; GCN-NEXT: v_or_b32_e32 v28, v43, v56
-; GCN-NEXT: v_or_b32_e32 v29, v44, v58
-; GCN-NEXT: v_or_b32_e32 v30, v46, v59
-; GCN-NEXT: v_or_b32_e32 v31, v47, v57
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x3000000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x3000000, v31
-; GCN-NEXT: .LBB29_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: .LBB29_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v31, 0x300
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v61
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v63
-; VI-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v59
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v62
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v58
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v60
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v54
-; VI-NEXT: v_add_u16_e32 v15, 3, v40
-; VI-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v14, v15
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v15, v16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v16, v16, v17
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v18
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v18, 0x300, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v18, v19
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v19, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 0x300, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v20, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v21, 0x300, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v21, v22
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v22, v23
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v23, v24
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v24, v25
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v25, v26
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v26, v27
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v27, v28
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v28, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v32, 0x300, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v33, 3, v33
-; VI-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: .LBB29_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB29_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: .LBB29_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB29_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v54
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v29, v29, v30
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v30, v30, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v31, v31, v32
-; GFX9-NEXT: .LBB29_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:592
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:588
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:584
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:468
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:456
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v136, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v137, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v138, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
-; GFX11-NEXT: v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
-; GFX11-NEXT: v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
-; GFX11-NEXT: v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
-; GFX11-NEXT: v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
-; GFX11-NEXT: v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
-; GFX11-NEXT: v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v114, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v115, off, s32
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v181, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v182, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v136, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v137, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v138, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v132, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v150, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v160, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v161, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v167, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v176, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v41, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v115
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v116
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v117
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v118
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v119
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v131
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v144
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v145
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v146
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v147
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v148
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v162
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v163
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v164
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v165
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v166
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v179, 8, v179
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v180, 8, v180
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v181
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v183, 8, v183
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v162, 8, v136
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v163, 8, v137
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v164, 8, v138
-; GFX11-NEXT: v_lshlrev_b16 v165, 8, v103
-; GFX11-NEXT: v_lshlrev_b16 v166, 8, v102
-; GFX11-NEXT: v_lshlrev_b16 v144, 8, v101
-; GFX11-NEXT: v_lshlrev_b16 v145, 8, v100
-; GFX11-NEXT: v_lshlrev_b16 v146, 8, v99
-; GFX11-NEXT: v_lshlrev_b16 v147, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v148, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v28
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v26
-; GFX11-NEXT: v_lshlrev_b16 v129, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v130, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v131, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v14
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v0
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v51
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v124
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v125
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v126
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v127
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v111
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v121
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v120
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v122
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v123
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v107
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v108
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v109
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v110
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v106
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v7, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v6, v12
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v92
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v78
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v77
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v76
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v75
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v59
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v93
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v94
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v95
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v104
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v105
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v79
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v88
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v89
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v90
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v91
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v58
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v44
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v43
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v41
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v40
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v176
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v167
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v61
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v62
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v63
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v72
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v73
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v45
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v46
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v47
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v56
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v57
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v161
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v160
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v150
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v132
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v113
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v179
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v180
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v181
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v182
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v183
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v162
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v163
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v164
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v165
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v166
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v84
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v82
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v144
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v145
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v146
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v147
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v148
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v119
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v128
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v129
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v130
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v131
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v66
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v64
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v114
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v115
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v116
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v117
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v118
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v99
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v100
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v101
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v102
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v103
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: .LBB29_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v51, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v5, v50, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v124, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v125, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v126, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v127, v3
-; GFX11-NEXT: v_add_nc_u16 v6, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v7, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_add_nc_u16 v8, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v36, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v10, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v2, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v11, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_or_b32_e32 v3, v111, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v120, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v121, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v6, v122, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v107, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v108, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v109, v10
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v38, 3
-; GFX11-NEXT: v_or_b32_e32 v11, v110, v11
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v106, v2
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_add_nc_u16 v7, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v32, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v9, v92, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v78, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v77, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v76, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v75, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v74, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v60, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v59, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v93, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v94, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v95, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v104, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v105, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v79, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v88, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v89, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v90, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v91, v16
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_add_nc_u16 v12, v58, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v44, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v43, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v41, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v40, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v178, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v177, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v176, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v167, 3
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v61, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v62, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v63, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v72, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v73, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v45, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v46, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v47, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v56, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v57, v21
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_add_nc_u16 v17, v161, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v160, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v150, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v149, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v133, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v132, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v113, 3
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v179, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v180, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v181, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v182, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v183, v21
-; GFX11-NEXT: v_or_b32_e32 v22, v162, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v163, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v164, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v165, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v166, v26
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_add_nc_u16 v22, v112, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v97, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v87, 3
-; GFX11-NEXT: v_add_nc_u16 v27, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v84, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v83, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v82, 3
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v144, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v145, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v146, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v147, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v148, v26
-; GFX11-NEXT: v_or_b32_e32 v27, v119, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v128, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v129, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v130, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v131, v31
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_add_nc_u16 v27, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v33, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v34, v66, 3
-; GFX11-NEXT: v_add_nc_u16 v35, v65, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v64, 3
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v114, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v115, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v116, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v117, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v118, v31
-; GFX11-NEXT: v_or_b32_e32 v32, v99, v32
-; GFX11-NEXT: v_or_b32_e32 v33, v100, v33
-; GFX11-NEXT: v_or_b32_e32 v34, v101, v34
-; GFX11-NEXT: v_or_b32_e32 v35, v102, v35
-; GFX11-NEXT: v_or_b32_e32 v36, v103, v36
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v34
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v36
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: .LBB29_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v138, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v137, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v136, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:516
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:580
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:584
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:588
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:592
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <64 x bfloat> @bitcast_v16i64_to_v64bf16(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; kill: killed $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v62
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: .LBB30_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_addc_u32_e32 v32, vcc, 0, v62, vcc
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v14
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v13
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v11
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v9
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v8
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v6
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v5
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v4
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v3
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v2
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v1
-; GCN-NEXT: .LBB30_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v61
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v59
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v56, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v47
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v46, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v45, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v44, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 20, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v33
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v17, 16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v18, v18, v19, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v20, v20, v21, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_alignbit_b32 v24, v24, v25, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_alignbit_b32 v26, v26, v27, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_alignbit_b32 v28, v28, v29, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_alignbit_b32 v30, v30, v31, 16
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_alignbit_b32 v32, v32, v33, 16
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_alignbit_b32 v34, v34, v35, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_alignbit_b32 v36, v36, v37, 16
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v38, v38, v39, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_alignbit_b32 v48, v48, v49, 16
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_alignbit_b32 v50, v50, v51, 16
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_alignbit_b32 v52, v52, v53, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_alignbit_b32 v54, v54, v55, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_alignbit_b32 v40, v40, v41, 16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v63
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_alignbit_b32 v42, v42, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v56, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB30_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: .LBB30_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB30_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: .LBB30_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB30_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: .LBB30_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <16 x i64> @bitcast_v64bf16_to_v16i64(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v33
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v42
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v51
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v63
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v60
-; GCN-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v61, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v58
-; GCN-NEXT: v_alignbit_b32 v2, v2, v59, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v56
-; GCN-NEXT: v_alignbit_b32 v3, v3, v57, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v46
-; GCN-NEXT: v_alignbit_b32 v4, v4, v47, 16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v45, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v11, v11, v12, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v13, v13, v14, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v34
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v33, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v17, v18, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v43
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v19, v19, v42, 16
-; GCN-NEXT: v_alignbit_b32 v20, v20, v44, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; GCN-NEXT: v_alignbit_b32 v22, v22, v48, 16
-; GCN-NEXT: v_alignbit_b32 v23, v23, v38, 16
-; GCN-NEXT: v_alignbit_b32 v24, v24, v50, 16
-; GCN-NEXT: v_alignbit_b32 v25, v25, v51, 16
-; GCN-NEXT: v_alignbit_b32 v26, v26, v53, 16
-; GCN-NEXT: v_alignbit_b32 v27, v27, v55, 16
-; GCN-NEXT: v_alignbit_b32 v28, v28, v40, 16
-; GCN-NEXT: v_alignbit_b32 v29, v29, v63, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; kill: killed $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: .LBB31_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v4, v5, v4, 16
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_alignbit_b32 v6, v7, v6, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v7, v8, v7, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v8, v9, v8, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_alignbit_b32 v10, v11, v10, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v11, v12, v11, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v12, v13, v12, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_alignbit_b32 v14, v15, v14, 16
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v34
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v15, v16, v15, 16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v33
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_alignbit_b32 v16, v17, v16, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v42
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v44
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v54
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v41
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v43
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v40, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v44, 0x40c00000, v31
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v32
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v33
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v18, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v19, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v20, v39, v54, 16
-; GCN-NEXT: v_alignbit_b32 v21, v48, v40, 16
-; GCN-NEXT: v_alignbit_b32 v22, v49, v22, 16
-; GCN-NEXT: v_alignbit_b32 v23, v50, v23, 16
-; GCN-NEXT: v_alignbit_b32 v24, v51, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v32, v25, 16
-; GCN-NEXT: v_alignbit_b32 v26, v33, v26, 16
-; GCN-NEXT: v_alignbit_b32 v27, v34, v27, 16
-; GCN-NEXT: v_alignbit_b32 v28, v35, v28, 16
-; GCN-NEXT: v_alignbit_b32 v29, v36, v29, 16
-; GCN-NEXT: v_alignbit_b32 v30, v37, v30, 16
-; GCN-NEXT: v_alignbit_b32 v31, v38, v31, 16
-; GCN-NEXT: .LBB31_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB31_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v15
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v14
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v13
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v12
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v11
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v10
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v9
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v8
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v7
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v6
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v5
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v4
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v3
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v2
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v1
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v0
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v31
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v30
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v29
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v29, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v28
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v28, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v27
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v27, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v26
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v26, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v25
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v25, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v24
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v24, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v23
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v23, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v22
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v22, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v21
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v20
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v20, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v19
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v19, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v18
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v17
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v17, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB31_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB31_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v32, s7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v31, v31, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v30, v30, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v29, v29, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v28, v28, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v27, v27, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v26, v26, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v25, v25, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v24, v24, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v23, v23, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v22, v22, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v21, v21, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v20, v20, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v19, v19, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v18, v18, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v17, v17, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v16, v16, v32, s7
-; GFX9-NEXT: .LBB31_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB31_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v32, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v32
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v39, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_add3_u32 v34, v34, v32, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v39, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v36, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v36, v36, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v32, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_perm_b32 v14, v14, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v33, v33, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v37, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v13, v13, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v34, v34, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v33, v33, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_bfe_u32 v35, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v35, v35, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v33, v33, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v35, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v33, v33, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GFX11-NEXT: v_perm_b32 v7, v7, v32, 0x7060302
-; GFX11-NEXT: v_perm_b32 v6, v6, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_bfe_u32 v34, v4, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v5, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v3
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v4
-; GFX11-NEXT: v_perm_b32 v5, v5, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v4, v4, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v35, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v2, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_perm_b32 v3, v3, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_perm_b32 v2, v2, v33, 0x7060302
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v34, v0, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v0
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_perm_b32 v1, v1, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v0, v0, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX11-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_bfe_u32 v35, v30, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v29
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v30, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_perm_b32 v31, v31, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v30, v30, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_add3_u32 v33, v33, v29, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v29
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-NEXT: v_bfe_u32 v34, v28, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v27
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v28, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v28
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v26
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_perm_b32 v29, v29, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v28, v28, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX11-NEXT: v_add3_u32 v33, v33, v27, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v27
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-NEXT: v_bfe_u32 v35, v26, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v25
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v26, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v26
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_perm_b32 v27, v27, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v26, v26, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-NEXT: v_add3_u32 v33, v33, v25, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v25
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-NEXT: v_bfe_u32 v34, v24, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v23
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v24
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_perm_b32 v25, v25, v32, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v24, v24, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v33, v23, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-NEXT: v_bfe_u32 v35, v22, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v22, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v22
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v23, v23, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v36, v34, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v21, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v21
-; GFX11-NEXT: v_perm_b32 v22, v22, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GFX11-NEXT: v_add3_u32 v35, v36, v34, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v21, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v33, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v32, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v32, v38, v33, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v21, v21, v34, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v34, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v35, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_add3_u32 v34, v34, v20, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v34, v35, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v36
-; GFX11-NEXT: v_add3_u32 v35, v37, v33, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v33
-; GFX11-NEXT: v_bfe_u32 v37, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_perm_b32 v20, v20, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v37, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v37, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_bfe_u32 v48, v18, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v36, v48, v18, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX11-NEXT: v_perm_b32 v19, v19, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v36, v37, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-NEXT: v_add3_u32 v37, v38, v35, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v35
-; GFX11-NEXT: v_bfe_u32 v39, v17, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_bfe_u32 v48, v36, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v36
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v16
-; GFX11-NEXT: v_perm_b32 v18, v18, v34, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v39, v17, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_bfe_u32 v37, v16, 16, 1
-; GFX11-NEXT: v_add3_u32 v48, v48, v36, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v38, v39, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_add3_u32 v37, v37, v16, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v17, v17, v35, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v48, v49, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v37, v50, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v16, v16, v36, 0x7060302
-; GFX11-NEXT: .LBB31_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <64 x half> @bitcast_v16i64_to_v64f16(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; kill: killed $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v62
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v62
-; GCN-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v63
-; GCN-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v37
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v38
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB32_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v32, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v34, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v36, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 3, v11
-; GCN-NEXT: v_addc_u32_e32 v38, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 3, v13
-; GCN-NEXT: v_addc_u32_e32 v48, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 3, v15
-; GCN-NEXT: v_addc_u32_e32 v50, vcc, 0, v16, vcc
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 3, v17
-; GCN-NEXT: v_addc_u32_e32 v52, vcc, 0, v18, vcc
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v63
-; GCN-NEXT: v_addc_u32_e32 v54, vcc, 0, v62, vcc
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v52
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v51
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v50
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v49
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v48
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v39
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v38
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v37
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v36
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v35
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v34
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v33
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v32
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v58
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v2
-; GCN-NEXT: .LBB32_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v60
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v59
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v58
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v2, v1
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v57
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v2, v1
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 12, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v2, v1
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v47
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 20, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v46
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v45
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v44
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v43
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v42
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v41
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v55
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v54
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v53
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v52
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v51
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v50
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v49
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v48
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v38
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v30, v34, v30
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x50, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v37
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x54, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_or_b32_e32 v35, v38, v35
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x58, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v33, v39, v33
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x5c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v32, v48, v32
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v49, v31
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v52, v53, v52
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v55, v54
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v40, v41, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v61, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v60, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB32_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: .LBB32_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB32_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: .LBB32_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB32_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: .LBB32_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <16 x i64> @bitcast_v64f16_to_v16i64(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v43
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v50
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v1
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v61
-; GCN-NEXT: v_or_b32_e32 v0, v62, v0
-; GCN-NEXT: v_or_b32_e32 v1, v60, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: v_or_b32_e32 v2, v58, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v57
-; GCN-NEXT: v_or_b32_e32 v3, v56, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v47
-; GCN-NEXT: v_or_b32_e32 v4, v46, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v32
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v37
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_or_b32_e32 v19, v43, v19
-; GCN-NEXT: v_or_b32_e32 v20, v41, v20
-; GCN-NEXT: v_or_b32_e32 v21, v55, v21
-; GCN-NEXT: v_or_b32_e32 v22, v49, v22
-; GCN-NEXT: v_or_b32_e32 v23, v50, v23
-; GCN-NEXT: v_or_b32_e32 v24, v39, v24
-; GCN-NEXT: v_or_b32_e32 v25, v36, v25
-; GCN-NEXT: v_or_b32_e32 v26, v48, v26
-; GCN-NEXT: v_or_b32_e32 v27, v52, v27
-; GCN-NEXT: v_or_b32_e32 v28, v53, v28
-; GCN-NEXT: v_or_b32_e32 v29, v54, v29
-; GCN-NEXT: v_or_b32_e32 v30, v40, v30
-; GCN-NEXT: v_or_b32_e32 v31, v42, v31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: .LBB33_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v41
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v55
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v50
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v39
-; GCN-NEXT: v_mov_b32_e32 v39, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v48
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v33
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_or_b32_e32 v19, v21, v20
-; GCN-NEXT: v_or_b32_e32 v20, v55, v39
-; GCN-NEXT: v_or_b32_e32 v21, v41, v48
-; GCN-NEXT: v_or_b32_e32 v22, v22, v49
-; GCN-NEXT: v_or_b32_e32 v23, v23, v50
-; GCN-NEXT: v_or_b32_e32 v24, v24, v51
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v38
-; GCN-NEXT: v_or_b32_e32 v28, v28, v33
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_or_b32_e32 v30, v30, v35
-; GCN-NEXT: v_or_b32_e32 v31, v31, v37
-; GCN-NEXT: .LBB33_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB33_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB33_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB33_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB33_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB33_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB33_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <64 x i16> @bitcast_v16i64_to_v64i16(<16 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i64_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; kill: killed $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; kill: killed $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v49, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v51, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v54, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v40, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v45, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB34_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; GCN-NEXT: v_alignbit_b32 v33, v31, v32, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v36, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v37, v24, v23, 16
-; GCN-NEXT: v_alignbit_b32 v38, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v39, v20, v19, 16
-; GCN-NEXT: v_alignbit_b32 v48, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v49, v16, v15, 16
-; GCN-NEXT: v_alignbit_b32 v51, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v54, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v40, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v43, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v45, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v57, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v59, v2, v1, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB34_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v59
-; GCN-NEXT: v_or_b32_e32 v1, v1, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v33
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v57
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v45
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v62
-; GCN-NEXT: v_or_b32_e32 v45, v1, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v43
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v40
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v60
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v54
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v58
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v51
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v56
-; GCN-NEXT: v_or_b32_e32 v14, v14, v33
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_or_b32_e32 v15, v15, v49
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v47
-; GCN-NEXT: v_or_b32_e32 v16, v16, v51
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v17, v17, v48
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v46
-; GCN-NEXT: v_or_b32_e32 v18, v18, v54
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_or_b32_e32 v19, v19, v39
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v20, v20, v34
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v38
-; GCN-NEXT: v_or_b32_e32 v21, v21, v34
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v42
-; GCN-NEXT: v_or_b32_e32 v22, v22, v34
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v37
-; GCN-NEXT: v_or_b32_e32 v23, v23, v34
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v41
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v36
-; GCN-NEXT: v_or_b32_e32 v25, v25, v34
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v55
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v35
-; GCN-NEXT: v_or_b32_e32 v27, v27, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v53
-; GCN-NEXT: v_or_b32_e32 v28, v28, v34
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v30, v30, v52
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_or_b32_e32 v32, v32, v43
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v31, v31, v50
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v44, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i64_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_addc_u32_e32 v31, vcc, 0, v31, vcc
-; VI-NEXT: v_add_u32_e32 v28, vcc, 3, v28
-; VI-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
-; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
-; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
-; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
-; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
-; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
-; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
-; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
-; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: .LBB34_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i64_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v30, vcc, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v28, vcc, 3, v28
-; GFX9-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
-; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
-; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
-; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
-; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: .LBB34_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i64_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v30, vcc_lo, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v28, vcc_lo, v28, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: .LBB34_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i64> %a, splat (i64 3)
- %a2 = bitcast <16 x i64> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i64> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v16i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v20
-; GCN-NEXT: v_mov_b32_e32 v38, v18
-; GCN-NEXT: v_mov_b32_e32 v39, v16
-; GCN-NEXT: v_mov_b32_e32 v48, v14
-; GCN-NEXT: v_mov_b32_e32 v49, v12
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v51, v8
-; GCN-NEXT: v_mov_b32_e32 v52, v6
-; GCN-NEXT: v_mov_b32_e32 v53, v4
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(12) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v24
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v55
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v36
-; GCN-NEXT: v_or_b32_e32 v1, v1, v58
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v53
-; GCN-NEXT: v_or_b32_e32 v2, v2, v57
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v52
-; GCN-NEXT: v_or_b32_e32 v3, v3, v35
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v4, v4, v60
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v50
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v49
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v48
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v39
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v38
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v43
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v46
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v40
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v62
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v47
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v44
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v18, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v21, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v29, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v30, v32
-; GCN-NEXT: v_or_b32_e32 v31, v31, v59
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: .LBB35_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v36, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v58, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v57, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v52
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v35, v3
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v60, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v32, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v32, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v15
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v32, v16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v32, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v32, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v32, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v32, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v32, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v32, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v32, v30
-; GCN-NEXT: v_or_b32_e32 v31, v59, v31
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x30000, v31
-; GCN-NEXT: .LBB35_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(11)
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v16i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v33, 3
-; VI-NEXT: v_add_u16_e32 v32, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v32, v15
-; VI-NEXT: v_add_u16_e32 v32, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v32, v14
-; VI-NEXT: v_add_u16_e32 v32, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v32, v13
-; VI-NEXT: v_add_u16_e32 v32, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v32, v12
-; VI-NEXT: v_add_u16_e32 v32, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v32, v11
-; VI-NEXT: v_add_u16_e32 v32, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v32, v10
-; VI-NEXT: v_add_u16_e32 v32, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v32, v9
-; VI-NEXT: v_add_u16_e32 v32, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v32, v8
-; VI-NEXT: v_add_u16_e32 v32, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v32, v7
-; VI-NEXT: v_add_u16_e32 v32, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v32, v6
-; VI-NEXT: v_add_u16_e32 v32, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v32, v5
-; VI-NEXT: v_add_u16_e32 v32, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v32, v4
-; VI-NEXT: v_add_u16_e32 v32, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v32, v3
-; VI-NEXT: v_add_u16_e32 v32, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v32, v2
-; VI-NEXT: v_add_u16_e32 v32, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v32, v1
-; VI-NEXT: v_add_u16_e32 v32, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v32, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v31
-; VI-NEXT: v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: v_add_u16_e32 v32, 3, v30
-; VI-NEXT: v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v32, v30
-; VI-NEXT: v_add_u16_e32 v32, 3, v29
-; VI-NEXT: v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v32, v29
-; VI-NEXT: v_add_u16_e32 v32, 3, v28
-; VI-NEXT: v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v32, v28
-; VI-NEXT: v_add_u16_e32 v32, 3, v27
-; VI-NEXT: v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_add_u16_e32 v32, 3, v26
-; VI-NEXT: v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v32, v26
-; VI-NEXT: v_add_u16_e32 v32, 3, v25
-; VI-NEXT: v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v32, v25
-; VI-NEXT: v_add_u16_e32 v32, 3, v24
-; VI-NEXT: v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v32, v24
-; VI-NEXT: v_add_u16_e32 v32, 3, v23
-; VI-NEXT: v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v32, v23
-; VI-NEXT: v_add_u16_e32 v32, 3, v22
-; VI-NEXT: v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v32, v22
-; VI-NEXT: v_add_u16_e32 v32, 3, v21
-; VI-NEXT: v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v32, v21
-; VI-NEXT: v_add_u16_e32 v32, 3, v20
-; VI-NEXT: v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v32, v20
-; VI-NEXT: v_add_u16_e32 v32, 3, v19
-; VI-NEXT: v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v32, v19
-; VI-NEXT: v_add_u16_e32 v32, 3, v18
-; VI-NEXT: v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v32, v18
-; VI-NEXT: v_add_u16_e32 v32, 3, v17
-; VI-NEXT: v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v32, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB35_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v16i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB35_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v16i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB35_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB35_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <16 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <16 x i64>
- br label %end
-
-end:
- %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i64> %phi
-}
-
-define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v51, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v52, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v53, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v61, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v62, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v63, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v2
-; GCN-NEXT: .LBB36_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GCN-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GCN-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GCN-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GCN-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GCN-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GCN-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GCN-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v51, v8, v7, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v6, v5, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v6, v5, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v52, v6, v5, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v4, v3, 24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v53, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v61, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v62, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v63, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 8, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 8, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v2
-; GCN-NEXT: .LBB36_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v63
-; GCN-NEXT: v_or_b32_e32 v1, v1, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v35
-; GCN-NEXT: v_or_b32_e32 v2, v2, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v61
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v33, v33, v34
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v35
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v33
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v63, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v62, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v61, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v27
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v29
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v30
-; GCN-NEXT: s_waitcnt vmcnt(2) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v53
-; GCN-NEXT: v_or_b32_e32 v48, v48, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v60
-; GCN-NEXT: v_or_b32_e32 v54, v54, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v52
-; GCN-NEXT: v_or_b32_e32 v60, v50, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v59
-; GCN-NEXT: v_or_b32_e32 v59, v49, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v51
-; GCN-NEXT: v_or_b32_e32 v32, v39, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v58
-; GCN-NEXT: v_or_b32_e32 v30, v38, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GCN-NEXT: v_or_b32_e32 v28, v37, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v57
-; GCN-NEXT: v_or_b32_e32 v26, v36, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GCN-NEXT: v_or_b32_e32 v24, v63, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v56
-; GCN-NEXT: v_or_b32_e32 v22, v62, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GCN-NEXT: v_or_b32_e32 v20, v61, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v47
-; GCN-NEXT: v_or_b32_e32 v16, v35, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: v_or_b32_e32 v15, v15, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v46
-; GCN-NEXT: v_or_b32_e32 v17, v34, v17
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_or_b32_e32 v19, v33, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v45
-; GCN-NEXT: v_or_b32_e32 v21, v18, v21
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GCN-NEXT: v_or_b32_e32 v23, v14, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v44
-; GCN-NEXT: v_or_b32_e32 v25, v13, v14
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GCN-NEXT: v_or_b32_e32 v27, v12, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v43
-; GCN-NEXT: v_or_b32_e32 v29, v11, v12
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GCN-NEXT: v_or_b32_e32 v31, v10, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v42
-; GCN-NEXT: v_or_b32_e32 v33, v9, v10
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GCN-NEXT: v_or_b32_e32 v35, v8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v41
-; GCN-NEXT: v_or_b32_e32 v52, v7, v8
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; GCN-NEXT: v_or_b32_e32 v50, v6, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v40
-; GCN-NEXT: v_or_b32_e32 v49, v5, v6
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GCN-NEXT: v_or_b32_e32 v51, v4, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v55
-; GCN-NEXT: v_or_b32_e32 v53, v3, v4
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v55, v2, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v41, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v10, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v12, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v5, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v6, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v7, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v8, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v9, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v11, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v13, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v18, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v34, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v36, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v37, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v38, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v39, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v40, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v42, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v43, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v44, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v45, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v46, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v47, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v56, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v57, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v58, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v61, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v62, v4, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v63, v4, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v3, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v54, v1, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v60
-; GCN-NEXT: v_or_b32_e32 v4, v2, v10
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v59
-; GCN-NEXT: v_or_b32_e32 v59, v10, v12
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v32
-; GCN-NEXT: v_or_b32_e32 v5, v10, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v7, v7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v11, v11, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v13, v13, v18
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v34
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v36
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v37
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v38
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v39
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v40
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v42
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v43
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v31, v31, v44
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GCN-NEXT: v_or_b32_e32 v33, v33, v45
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GCN-NEXT: v_or_b32_e32 v35, v35, v46
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v52
-; GCN-NEXT: v_or_b32_e32 v37, v37, v47
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff, v50
-; GCN-NEXT: v_or_b32_e32 v39, v39, v56
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff, v49
-; GCN-NEXT: v_or_b32_e32 v49, v49, v57
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v51, v51, v58
-; GCN-NEXT: v_add_i32_e32 v45, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v53
-; GCN-NEXT: v_or_b32_e32 v53, v53, v61
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GCN-NEXT: v_or_b32_e32 v55, v55, v62
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: v_or_b32_e32 v41, v41, v63
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v59, v60, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v5, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v8, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v30, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v44, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v45, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v51, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v53, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v55, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; kill: killed $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v56, 24, v28
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v60, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v61, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v52, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v53, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v42, 24, v32
-; VI-NEXT: .LBB36_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; VI-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; VI-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; VI-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; VI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; VI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; VI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; VI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; VI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; VI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; VI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; VI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; VI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; VI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; VI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; VI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[39:40], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b32_e32 v42, 24, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b32_e32 v56, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v57, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v60, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v61, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v52, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v53, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v1
-; VI-NEXT: .LBB36_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v55
-; VI-NEXT: v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v41
-; VI-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v54, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; VI-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v53
-; VI-NEXT: v_or_b32_sdwa v1, v45, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v44
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v52
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v38
-; VI-NEXT: v_or_b32_sdwa v1, v43, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v63
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v51
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v33
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v36, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v49
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v56
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v42
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; kill: killed $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; kill: killed $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(32)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_waitcnt vmcnt(31)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[3:4]
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: .LBB36_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GFX9-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GFX9-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GFX9-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GFX9-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GFX9-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GFX9-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GFX9-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; GFX9-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX9-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[1:2]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v18
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v17
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v1
-; GFX9-NEXT: .LBB36_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v40
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v43
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v63
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v53
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v61
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v48
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v46
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v41
-; GFX9-NEXT: v_or_b32_sdwa v34, v54, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v62, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: .LBB36_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GFX11-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GFX11-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v1
-; GFX11-NEXT: .LBB36_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v63
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v65
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v61
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v73
-; GFX11-NEXT: v_or_b32_e32 v65, v67, v65
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v66
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v62
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v66
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v60
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v65
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v67
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v56
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v65
-; GFX11-NEXT: v_or_b32_e32 v55, v66, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v47
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v45
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v64
-; GFX11-NEXT: v_or_b32_e32 v55, v65, v66
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v44
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v43
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v54, v65, v54
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v40
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v55
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v65
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v182
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v181
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v55
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v65
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v54
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v53, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v176
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v167
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v53, v53, v54
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v52
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v65
-; GFX11-NEXT: v_or_b32_e32 v54, v66, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v53
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v54
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v164
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v160
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v150
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v146
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v49
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v145
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v144
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v135
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v132
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v131
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v128
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v119
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v117
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v116
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v115
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v103
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v102
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v101
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v99
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v86
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v84
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v83
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v81
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v70
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v69
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v68
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v10
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:148
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:160
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:192
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:212
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:244
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:288
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:340
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:352
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:380
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:376
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v1
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:372
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:364
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v0, v39
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v47
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v54
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v46
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v37
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v48
-; GCN-NEXT: v_or_b32_e32 v8, v8, v45
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v53
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v42
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v41
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v40
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v63
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v50
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v51
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v49
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v20, v60
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v58
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v22, v62
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v59
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v61
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v34
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v52
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v35
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v33
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v44
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v57
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v33, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v38, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v48, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v55, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v54, v49
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v18, v18, v50
-; GCN-NEXT: v_or_b32_e32 v19, v19, v51
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_or_b32_e32 v21, v21, v53
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: v_or_b32_e32 v23, v23, v33
-; GCN-NEXT: v_or_b32_e32 v24, v24, v34
-; GCN-NEXT: v_or_b32_e32 v25, v25, v35
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v37
-; GCN-NEXT: v_or_b32_e32 v28, v28, v38
-; GCN-NEXT: v_or_b32_e32 v29, v29, v39
-; GCN-NEXT: v_or_b32_e32 v30, v30, v48
-; GCN-NEXT: v_or_b32_e32 v31, v31, v49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; kill: killed $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB37_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v39, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v56, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v47, v3
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v54, v4
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v46, v5
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v37, v6
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v43, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v45, v8
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v42, v10
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v41, v11
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v40, v12
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v63, v13
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v50, v14
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v0, v15
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v0, v16
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v0, v17
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v51, v18
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v49, v19
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v60, v20
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v58, v21
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v25, v62, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v29, v59, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v37, v32, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v50, v61, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v41, v34, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v45, v52, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v56, v35, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v58, v33, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v59, v44, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v57
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v57, v36, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v23, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v26, v24
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v48, v35
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v48
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v53, v48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_mov_b32_e32 v0, v55
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v55, v53
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v40, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v40
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v42, v40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v44, v43
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v46, v44
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v46
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_or_b32_e32 v46, v0, v46
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v47
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v0, v47
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v60, v0
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v61, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v62, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v63, v3
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s7, v37
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s7, v50
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s7, v41
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s7, v45
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s7, v56
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 0x300, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff, v56
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_or_b32_e32 v4, v36, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v39, v6
-; GCN-NEXT: v_or_b32_e32 v7, v49, v7
-; GCN-NEXT: v_or_b32_e32 v8, v51, v8
-; GCN-NEXT: v_or_b32_e32 v9, v52, v9
-; GCN-NEXT: v_or_b32_e32 v10, v54, v10
-; GCN-NEXT: v_or_b32_e32 v11, v22, v11
-; GCN-NEXT: v_or_b32_e32 v12, v23, v12
-; GCN-NEXT: v_or_b32_e32 v13, v24, v13
-; GCN-NEXT: v_or_b32_e32 v14, v26, v14
-; GCN-NEXT: v_or_b32_e32 v15, v27, v15
-; GCN-NEXT: v_or_b32_e32 v16, v28, v16
-; GCN-NEXT: v_or_b32_e32 v17, v30, v17
-; GCN-NEXT: v_or_b32_e32 v18, v31, v18
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: v_or_b32_e32 v20, v33, v20
-; GCN-NEXT: v_or_b32_e32 v21, v34, v21
-; GCN-NEXT: v_or_b32_e32 v22, v35, v25
-; GCN-NEXT: v_or_b32_e32 v23, v48, v29
-; GCN-NEXT: v_or_b32_e32 v24, v53, v37
-; GCN-NEXT: v_or_b32_e32 v25, v55, v50
-; GCN-NEXT: v_or_b32_e32 v26, v40, v41
-; GCN-NEXT: v_or_b32_e32 v27, v42, v45
-; GCN-NEXT: v_or_b32_e32 v28, v43, v56
-; GCN-NEXT: v_or_b32_e32 v29, v44, v58
-; GCN-NEXT: v_or_b32_e32 v30, v46, v59
-; GCN-NEXT: v_or_b32_e32 v31, v47, v57
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x3000000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x3000000, v31
-; GCN-NEXT: .LBB37_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB37_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: .LBB37_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB37_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v31, 0x300
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v61
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v63
-; VI-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v59
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v62
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v58
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v60
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v54
-; VI-NEXT: v_add_u16_e32 v15, 3, v40
-; VI-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v14, v15
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v15, v16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v16, v16, v17
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v18
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v18, 0x300, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v18, v19
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v19
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v19, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 0x300, v20
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v20, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v21, 0x300, v21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v21, v22
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v22, v23
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v23, v24
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v24, v25
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v25, v26
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v26, v27
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v27, v28
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v28, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v32, 3, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v32, 0x300, v32
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v33, 3, v33
-; VI-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: .LBB37_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v40
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v45
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v46
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v47
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v56
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v57
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB37_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v40, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: .LBB37_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB37_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v49, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v39, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v37, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v54
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT: v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v19
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v23
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v28, 3, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v29, 3, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v29
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v29, v29, v30
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v30, 3, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v30, v30, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v31, v31, v32
-; GFX9-NEXT: .LBB37_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:592
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:588
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:584
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:468
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:456
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v136, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v137, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v138, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
-; GFX11-NEXT: v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
-; GFX11-NEXT: v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
-; GFX11-NEXT: v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
-; GFX11-NEXT: v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
-; GFX11-NEXT: v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
-; GFX11-NEXT: v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v114, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v115, off, s32
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v181, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v182, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v136, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v137, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v138, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v132, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v150, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v160, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v161, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v167, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v176, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v41, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v115
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v116
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v117
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v118
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v119
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v131
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v144
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v145
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v146
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v147
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v148
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v162
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v163
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v164
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v165
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v166
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v179, 8, v179
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v180, 8, v180
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v181
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v183, 8, v183
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v162, 8, v136
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v163, 8, v137
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v164, 8, v138
-; GFX11-NEXT: v_lshlrev_b16 v165, 8, v103
-; GFX11-NEXT: v_lshlrev_b16 v166, 8, v102
-; GFX11-NEXT: v_lshlrev_b16 v144, 8, v101
-; GFX11-NEXT: v_lshlrev_b16 v145, 8, v100
-; GFX11-NEXT: v_lshlrev_b16 v146, 8, v99
-; GFX11-NEXT: v_lshlrev_b16 v147, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v148, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v28
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v26
-; GFX11-NEXT: v_lshlrev_b16 v129, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v130, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v131, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v14
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v0
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB37_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v51
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v124
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v125
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v126
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v127
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v111
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v121
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v120
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v122
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v123
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v107
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v108
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v109
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v110
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v106
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v7, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v6, v12
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v92
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v78
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v77
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v76
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v75
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v74
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v59
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v93
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v94
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v95
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v104
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v105
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v79
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v88
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v89
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v90
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v91
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v58
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v44
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v43
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v41
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v40
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v176
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v167
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v61
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v62
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v63
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v72
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v73
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v45
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v46
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v47
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v56
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v57
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v161
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v160
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v150
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v132
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v113
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v179
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v180
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v181
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v182
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v183
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v162
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v163
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v164
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v165
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v166
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v84
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v82
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v144
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v145
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v146
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v147
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v148
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v119
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v128
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v129
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v130
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v131
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v66
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v64
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v114
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v115
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v116
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v117
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v118
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v99
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v100
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v101
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v102
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v103
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: .LBB37_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB37_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v51, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v5, v50, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v124, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v125, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v126, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v127, v3
-; GFX11-NEXT: v_add_nc_u16 v6, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v7, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_add_nc_u16 v8, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v36, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v10, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v2, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v11, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_or_b32_e32 v3, v111, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v120, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v121, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v6, v122, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v107, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v108, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v109, v10
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v38, 3
-; GFX11-NEXT: v_or_b32_e32 v11, v110, v11
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v106, v2
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v6, v10, v11
-; GFX11-NEXT: v_add_nc_u16 v7, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v32, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v9, v92, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v78, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v77, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v76, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v75, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v74, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v60, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v59, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v93, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v94, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v95, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v104, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v105, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v79, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v88, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v89, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v90, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v91, v16
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v16
-; GFX11-NEXT: v_add_nc_u16 v12, v58, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v44, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v43, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v41, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v40, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v178, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v177, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v176, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v167, 3
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v61, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v62, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v63, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v72, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v73, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v45, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v46, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v47, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v56, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v57, v21
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v14, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-NEXT: v_add_nc_u16 v17, v161, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v160, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v150, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v149, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v133, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v132, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v113, 3
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v179, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v180, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v181, v19
-; GFX11-NEXT: v_or_b32_e32 v20, v182, v20
-; GFX11-NEXT: v_or_b32_e32 v21, v183, v21
-; GFX11-NEXT: v_or_b32_e32 v22, v162, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v163, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v164, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v165, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v166, v26
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_add_nc_u16 v22, v112, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v97, 3
-; GFX11-NEXT: v_add_nc_u16 v25, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v26, v87, 3
-; GFX11-NEXT: v_add_nc_u16 v27, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v84, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v83, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v82, 3
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v144, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v145, v23
-; GFX11-NEXT: v_or_b32_e32 v24, v146, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v147, v25
-; GFX11-NEXT: v_or_b32_e32 v26, v148, v26
-; GFX11-NEXT: v_or_b32_e32 v27, v119, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v128, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v129, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v130, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v131, v31
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v24
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v25
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v26
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v23, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v24, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v25, v28, v29
-; GFX11-NEXT: v_or_b32_e32 v26, v30, v31
-; GFX11-NEXT: v_add_nc_u16 v27, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v28, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v29, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v30, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v31, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v33, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v34, v66, 3
-; GFX11-NEXT: v_add_nc_u16 v35, v65, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v64, 3
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v114, v27
-; GFX11-NEXT: v_or_b32_e32 v28, v115, v28
-; GFX11-NEXT: v_or_b32_e32 v29, v116, v29
-; GFX11-NEXT: v_or_b32_e32 v30, v117, v30
-; GFX11-NEXT: v_or_b32_e32 v31, v118, v31
-; GFX11-NEXT: v_or_b32_e32 v32, v99, v32
-; GFX11-NEXT: v_or_b32_e32 v33, v100, v33
-; GFX11-NEXT: v_or_b32_e32 v34, v101, v34
-; GFX11-NEXT: v_or_b32_e32 v35, v102, v35
-; GFX11-NEXT: v_or_b32_e32 v36, v103, v36
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v27
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v28
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v29
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v30
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v31
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v34
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v36
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v28, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v29, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v30, v33, v34
-; GFX11-NEXT: v_or_b32_e32 v31, v35, v36
-; GFX11-NEXT: .LBB37_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v138, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v137, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v136, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:516
-; GFX11-NEXT: s_clause 0x12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:580
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:584
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:588
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:592
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <64 x bfloat> @bitcast_v16f64_to_v64bf16(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v15
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v14
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v13
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v12
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v11
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v10
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v9
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v8
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v7
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v5
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v4
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v3
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v2
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: .LBB38_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GCN-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GCN-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GCN-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GCN-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GCN-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GCN-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GCN-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v15
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v14
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v13
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v12
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v11
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v10
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v9
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v8
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v7
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v5
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v4
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v3
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v2
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v1
-; GCN-NEXT: .LBB38_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v34
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v33
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v61
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v60, v1, v2, 16
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v59
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v58, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v57
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v56, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 12, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v47
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v46, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 16, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 20, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v17, 16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v18, v18, v19, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v20, v20, v21, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_alignbit_b32 v24, v24, v25, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_alignbit_b32 v26, v26, v27, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_alignbit_b32 v28, v28, v29, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_alignbit_b32 v30, v30, v31, 16
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_alignbit_b32 v32, v32, v33, 16
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_alignbit_b32 v34, v34, v35, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_alignbit_b32 v36, v36, v37, 16
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v38, v38, v39, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_alignbit_b32 v48, v48, v49, 16
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_alignbit_b32 v50, v50, v51, 16
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_alignbit_b32 v52, v52, v53, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_alignbit_b32 v54, v54, v55, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_alignbit_b32 v40, v40, v41, 16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_alignbit_b32 v42, v42, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v60, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v44, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB38_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: .LBB38_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB38_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: .LBB38_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB38_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: .LBB38_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <16 x double> @bitcast_v64bf16_to_v16f64(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v33
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v42
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v51
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v63
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v0
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v60
-; GCN-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v61, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v58
-; GCN-NEXT: v_alignbit_b32 v2, v2, v59, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v56
-; GCN-NEXT: v_alignbit_b32 v3, v3, v57, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v46
-; GCN-NEXT: v_alignbit_b32 v4, v4, v47, 16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v45, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v11, v11, v12, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v13, v13, v14, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v14, v14, v15, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v34
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v33, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v17, v18, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v43
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v19, v19, v42, 16
-; GCN-NEXT: v_alignbit_b32 v20, v20, v44, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; GCN-NEXT: v_alignbit_b32 v22, v22, v48, 16
-; GCN-NEXT: v_alignbit_b32 v23, v23, v38, 16
-; GCN-NEXT: v_alignbit_b32 v24, v24, v50, 16
-; GCN-NEXT: v_alignbit_b32 v25, v25, v51, 16
-; GCN-NEXT: v_alignbit_b32 v26, v26, v53, 16
-; GCN-NEXT: v_alignbit_b32 v27, v27, v55, 16
-; GCN-NEXT: v_alignbit_b32 v28, v28, v40, 16
-; GCN-NEXT: v_alignbit_b32 v29, v29, v63, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; kill: killed $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: .LBB39_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v4, v5, v4, 16
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_alignbit_b32 v6, v7, v6, 16
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v7, v8, v7, 16
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v8, v9, v8, 16
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_alignbit_b32 v10, v11, v10, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v11, v12, v11, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v12, v13, v12, 16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_alignbit_b32 v14, v15, v14, 16
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v34
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v15, v16, v15, 16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v33
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_alignbit_b32 v16, v17, v16, 16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v42
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v44
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v54
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v41
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v43
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v40, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v44, 0x40c00000, v31
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v32
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v33
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_alignbit_b32 v18, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v19, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v20, v39, v54, 16
-; GCN-NEXT: v_alignbit_b32 v21, v48, v40, 16
-; GCN-NEXT: v_alignbit_b32 v22, v49, v22, 16
-; GCN-NEXT: v_alignbit_b32 v23, v50, v23, 16
-; GCN-NEXT: v_alignbit_b32 v24, v51, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v32, v25, 16
-; GCN-NEXT: v_alignbit_b32 v26, v33, v26, 16
-; GCN-NEXT: v_alignbit_b32 v27, v34, v27, 16
-; GCN-NEXT: v_alignbit_b32 v28, v35, v28, 16
-; GCN-NEXT: v_alignbit_b32 v29, v36, v29, 16
-; GCN-NEXT: v_alignbit_b32 v30, v37, v30, 16
-; GCN-NEXT: v_alignbit_b32 v31, v38, v31, 16
-; GCN-NEXT: .LBB39_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB39_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v15
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v14
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v13
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v12
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v11
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v10
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v9
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v8
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v7
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v6
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v5
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v4
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v3
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v2
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v1
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v0
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v32, 16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v31
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v31, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v30
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v30, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v29
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v29, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v28
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v28, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v27
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v27, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v26
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v26, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v25
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v25, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v24
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v24, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v23
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v23, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v22
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v22, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v21
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v21, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v20
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v20, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v19
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v19, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v18
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v18, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v17
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v17, v32, 16
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB39_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB39_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v33, v34, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v32, s7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v31, v31, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v30, v30, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v29, v29, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v28, v28, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v27, v27, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v26, v26, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v25, v25, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v24, v24, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v23, v23, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v22, v22, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v21, v21, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v20, v20, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v19, v19, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v18, v18, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v17, v17, v32, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_perm_b32 v16, v16, v32, s7
-; GFX9-NEXT: .LBB39_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB39_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v32, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v32
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v39, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_add3_u32 v34, v34, v32, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v39, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v36, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v36, v36, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v32, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_perm_b32 v14, v14, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v33, v33, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v37, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v13, v13, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v34, v34, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v33, v33, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_bfe_u32 v35, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v35, v35, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v33, v33, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v32, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v35, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v33, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v33, v33, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GFX11-NEXT: v_perm_b32 v7, v7, v32, 0x7060302
-; GFX11-NEXT: v_perm_b32 v6, v6, v33, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_bfe_u32 v34, v4, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v33, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v5, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v3
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v4
-; GFX11-NEXT: v_perm_b32 v5, v5, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v4, v4, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v35, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v2, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_bfe_u32 v33, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_perm_b32 v3, v3, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_perm_b32 v2, v2, v33, 0x7060302
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v34, v0, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_bfe_u32 v33, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v33, v33, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v0
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v30
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_perm_b32 v1, v1, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v0, v0, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v31, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX11-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_bfe_u32 v35, v30, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v29
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v30, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_perm_b32 v31, v31, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v30, v30, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v29, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_add3_u32 v33, v33, v29, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v29
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-NEXT: v_bfe_u32 v34, v28, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v27
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v28, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v28
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v26
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_perm_b32 v29, v29, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_perm_b32 v28, v28, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v27, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX11-NEXT: v_add3_u32 v33, v33, v27, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v27
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-NEXT: v_bfe_u32 v35, v26, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v25
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v26, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v26
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_perm_b32 v27, v27, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v32, v34, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v26, v26, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v25, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v34, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-NEXT: v_add3_u32 v33, v33, v25, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v25
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-NEXT: v_bfe_u32 v34, v24, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v35, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v23
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v24
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_perm_b32 v25, v25, v32, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX11-NEXT: v_bfe_u32 v32, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v34, v38, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_add3_u32 v32, v32, v35, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_perm_b32 v24, v24, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v33, v23, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v33, v23, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-NEXT: v_bfe_u32 v35, v22, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v33, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v33, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v22, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v22
-; GFX11-NEXT: v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v23, v23, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v36, v34, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v32, v21, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v21
-; GFX11-NEXT: v_perm_b32 v22, v22, v33, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v33, 0x40c00000, v35
-; GFX11-NEXT: v_add3_u32 v35, v36, v34, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v32, v32, v21, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v33, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v32, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v32, v38, v33, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v21, v21, v34, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v34, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v35, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_add3_u32 v34, v34, v20, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v34, v35, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v36
-; GFX11-NEXT: v_add3_u32 v35, v37, v33, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v33
-; GFX11-NEXT: v_bfe_u32 v37, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_perm_b32 v20, v20, v32, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v37, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v37, v38, v34, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_bfe_u32 v48, v18, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v36, v48, v18, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX11-NEXT: v_perm_b32 v19, v19, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v35, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v36, v37, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-NEXT: v_add3_u32 v37, v38, v35, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v35
-; GFX11-NEXT: v_bfe_u32 v39, v17, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_bfe_u32 v48, v36, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v36
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v16
-; GFX11-NEXT: v_perm_b32 v18, v18, v34, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v39, v17, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_bfe_u32 v37, v16, 16, 1
-; GFX11-NEXT: v_add3_u32 v48, v48, v36, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v38, v39, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_add3_u32 v37, v37, v16, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v17, v17, v35, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v48, v49, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v37, v50, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v16, v16, v36, 0x7060302
-; GFX11-NEXT: .LBB39_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <64 x half> @bitcast_v16f64_to_v64f16(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; kill: killed $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; kill: killed $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v13
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v12
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v11
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v10
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v8
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v7
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v49
-; GCN-NEXT: v_mov_b32_e32 v49, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v62, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v63, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: .LBB40_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GCN-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GCN-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GCN-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GCN-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GCN-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GCN-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GCN-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v7
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v24
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v25
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v26
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v27
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v3
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v62, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v63, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v3
-; GCN-NEXT: .LBB40_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v63, v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v62
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v2, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v61
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v61, v3, v1
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 12, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v60
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v60, v4, v3
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v59
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v58
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v58, v4, v3
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v57
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v56
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v47
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v45
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v44
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v42
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v41
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v40
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v53
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v52
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v51
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v50
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v48
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v38
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v37
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v35
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v39
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v49
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v49, v48
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v52, v53, v52
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v55, v54
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v40, v41, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v63, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v60, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB40_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: .LBB40_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB40_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: .LBB40_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB40_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: .LBB40_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <16 x double> @bitcast_v64f16_to_v16f64(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:116
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v43
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v55
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v50
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v1
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v63
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v61
-; GCN-NEXT: v_or_b32_e32 v0, v62, v0
-; GCN-NEXT: v_or_b32_e32 v1, v60, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: v_or_b32_e32 v2, v58, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v57
-; GCN-NEXT: v_or_b32_e32 v3, v56, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v47
-; GCN-NEXT: v_or_b32_e32 v4, v46, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v32
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v37
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_or_b32_e32 v19, v43, v19
-; GCN-NEXT: v_or_b32_e32 v20, v41, v20
-; GCN-NEXT: v_or_b32_e32 v21, v55, v21
-; GCN-NEXT: v_or_b32_e32 v22, v49, v22
-; GCN-NEXT: v_or_b32_e32 v23, v50, v23
-; GCN-NEXT: v_or_b32_e32 v24, v39, v24
-; GCN-NEXT: v_or_b32_e32 v25, v36, v25
-; GCN-NEXT: v_or_b32_e32 v26, v48, v26
-; GCN-NEXT: v_or_b32_e32 v27, v52, v27
-; GCN-NEXT: v_or_b32_e32 v28, v53, v28
-; GCN-NEXT: v_or_b32_e32 v29, v54, v29
-; GCN-NEXT: v_or_b32_e32 v30, v40, v30
-; GCN-NEXT: v_or_b32_e32 v31, v42, v31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: .LBB41_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v58
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v56
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v46
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v45
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v6, v5
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v11, v12, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v51
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v17, v18, v17
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v41
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v55
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v50
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v39
-; GCN-NEXT: v_mov_b32_e32 v39, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v48
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v42
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v33
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_or_b32_e32 v19, v21, v20
-; GCN-NEXT: v_or_b32_e32 v20, v55, v39
-; GCN-NEXT: v_or_b32_e32 v21, v41, v48
-; GCN-NEXT: v_or_b32_e32 v22, v22, v49
-; GCN-NEXT: v_or_b32_e32 v23, v23, v50
-; GCN-NEXT: v_or_b32_e32 v24, v24, v51
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: v_or_b32_e32 v26, v26, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v38
-; GCN-NEXT: v_or_b32_e32 v28, v28, v33
-; GCN-NEXT: v_or_b32_e32 v29, v29, v34
-; GCN-NEXT: v_or_b32_e32 v30, v30, v35
-; GCN-NEXT: v_or_b32_e32 v31, v31, v37
-; GCN-NEXT: .LBB41_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB41_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB41_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB41_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB41_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB41_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB41_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <64 x i16> @bitcast_v16f64_to_v64i16(<16 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f64_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; kill: killed $vgpr34
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v47, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v45, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v63, v2, v1, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB42_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GCN-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
-; GCN-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
-; GCN-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
-; GCN-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
-; GCN-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
-; GCN-NEXT: v_add_f64 v[27:28], v[27:28], 1.0
-; GCN-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f64 v[31:32], v[31:32], 1.0
-; GCN-NEXT: v_alignbit_b32 v33, v32, v31, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v30, v29, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v28, v27, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v26, v25, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v24, v23, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v20, v19, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v12, v11, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v10, v9, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v47, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v45, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v63, v2, v1, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GCN-NEXT: .LBB42_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v63
-; GCN-NEXT: v_or_b32_e32 v1, v1, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v33
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v42, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v7
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v30
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v45
-; GCN-NEXT: v_or_b32_e32 v46, v44, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v62
-; GCN-NEXT: v_or_b32_e32 v62, v42, v6
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 8, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v47
-; GCN-NEXT: v_or_b32_e32 v44, v41, v6
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 12, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v61
-; GCN-NEXT: v_or_b32_e32 v45, v4, v6
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v1, v1, v4
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 20, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v60
-; GCN-NEXT: v_or_b32_e32 v2, v2, v4
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v3, v3, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 28, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v59
-; GCN-NEXT: v_or_b32_e32 v5, v5, v8
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v7, v7, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 36, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v58
-; GCN-NEXT: v_or_b32_e32 v9, v9, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v11, v11, v14
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 44, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v43
-; GCN-NEXT: v_or_b32_e32 v13, v13, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v15, v15, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v17, v40, v17
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_or_b32_e32 v19, v55, v19
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v21, v53, v21
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v51, v23
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v25, v36, v25
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_or_b32_e32 v27, v49, v27
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_or_b32_e32 v29, v38, v29
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v37, v31
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v33, v33, v36
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v48, v36
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_or_b32_e32 v39, v39, v49
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_or_b32_e32 v35, v35, v51
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_or_b32_e32 v50, v50, v53
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_or_b32_e32 v34, v34, v55
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v54, v54, v40
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_or_b32_e32 v52, v52, v41
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_or_b32_e32 v41, v63, v41
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v46, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v62, v56, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v47, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v61, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v30, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f64_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; VI-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: .LBB42_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f64_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX9-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: .LBB42_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f64_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-NEXT: v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: .LBB42_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <16 x double> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x double> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v16f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v20
-; GCN-NEXT: v_mov_b32_e32 v38, v18
-; GCN-NEXT: v_mov_b32_e32 v39, v16
-; GCN-NEXT: v_mov_b32_e32 v48, v14
-; GCN-NEXT: v_mov_b32_e32 v49, v12
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v51, v8
-; GCN-NEXT: v_mov_b32_e32 v52, v6
-; GCN-NEXT: v_mov_b32_e32 v53, v4
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(12) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v24
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v55
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v36
-; GCN-NEXT: v_or_b32_e32 v1, v1, v58
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v53
-; GCN-NEXT: v_or_b32_e32 v2, v2, v57
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v52
-; GCN-NEXT: v_or_b32_e32 v3, v3, v35
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v4, v4, v60
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v50
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v49
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v48
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v39
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v38
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v43
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v56
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v46
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v40
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v62
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v47
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v44
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v18, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v21, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v29, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v30, v32
-; GCN-NEXT: v_or_b32_e32 v31, v31, v59
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; kill: killed $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: .LBB43_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v36, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v58, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v57, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v52
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v35, v3
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v61
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v4, v60, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v32, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v32, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v15
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v32, v16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v32, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v32, v19
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v32, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v32, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v32, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v32, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v32, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v32, v30
-; GCN-NEXT: v_or_b32_e32 v31, v59, v31
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x30000, v31
-; GCN-NEXT: .LBB43_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(11)
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v16f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB43_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v33, 3
-; VI-NEXT: v_add_u16_e32 v32, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v32, v15
-; VI-NEXT: v_add_u16_e32 v32, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v32, v14
-; VI-NEXT: v_add_u16_e32 v32, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v32, v13
-; VI-NEXT: v_add_u16_e32 v32, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v32, v12
-; VI-NEXT: v_add_u16_e32 v32, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v32, v11
-; VI-NEXT: v_add_u16_e32 v32, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v32, v10
-; VI-NEXT: v_add_u16_e32 v32, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v32, v9
-; VI-NEXT: v_add_u16_e32 v32, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v32, v8
-; VI-NEXT: v_add_u16_e32 v32, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v32, v7
-; VI-NEXT: v_add_u16_e32 v32, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v32, v6
-; VI-NEXT: v_add_u16_e32 v32, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v32, v5
-; VI-NEXT: v_add_u16_e32 v32, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v32, v4
-; VI-NEXT: v_add_u16_e32 v32, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v32, v3
-; VI-NEXT: v_add_u16_e32 v32, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v32, v2
-; VI-NEXT: v_add_u16_e32 v32, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v32, v1
-; VI-NEXT: v_add_u16_e32 v32, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v32, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v32, 3, v31
-; VI-NEXT: v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v32, v31
-; VI-NEXT: v_add_u16_e32 v32, 3, v30
-; VI-NEXT: v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v32, v30
-; VI-NEXT: v_add_u16_e32 v32, 3, v29
-; VI-NEXT: v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v32, v29
-; VI-NEXT: v_add_u16_e32 v32, 3, v28
-; VI-NEXT: v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v32, v28
-; VI-NEXT: v_add_u16_e32 v32, 3, v27
-; VI-NEXT: v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_add_u16_e32 v32, 3, v26
-; VI-NEXT: v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v32, v26
-; VI-NEXT: v_add_u16_e32 v32, 3, v25
-; VI-NEXT: v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v32, v25
-; VI-NEXT: v_add_u16_e32 v32, 3, v24
-; VI-NEXT: v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v32, v24
-; VI-NEXT: v_add_u16_e32 v32, 3, v23
-; VI-NEXT: v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v32, v23
-; VI-NEXT: v_add_u16_e32 v32, 3, v22
-; VI-NEXT: v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v32, v22
-; VI-NEXT: v_add_u16_e32 v32, 3, v21
-; VI-NEXT: v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v32, v21
-; VI-NEXT: v_add_u16_e32 v32, 3, v20
-; VI-NEXT: v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v32, v20
-; VI-NEXT: v_add_u16_e32 v32, 3, v19
-; VI-NEXT: v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v32, v19
-; VI-NEXT: v_add_u16_e32 v32, 3, v18
-; VI-NEXT: v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v32, v18
-; VI-NEXT: v_add_u16_e32 v32, 3, v17
-; VI-NEXT: v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v32, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB43_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v16f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB43_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB43_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v16f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB43_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB43_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <16 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <16 x double>
- br label %end
-
-end:
- %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x double> %phi
-}
-
-define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:64
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:56
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:48
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v6
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v12
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v14
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v16
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v20
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v24
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v28
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v30
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:392
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v11
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v10
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v9
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v7
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v4
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:128
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:144
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:136
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:160
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:176
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:168
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:208
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:240
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:236
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:256
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:272
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:268
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:264
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:304
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:300
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:296
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:332
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:328
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:344
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:376
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:372
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:364
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:360
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v55, v1, v15
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v41, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v45, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GCN-NEXT: v_or_b32_e32 v22, v1, v22
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v1, v52
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v56
-; GCN-NEXT: v_or_b32_e32 v5, v1, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v31
-; GCN-NEXT: v_or_b32_e32 v13, v1, v13
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v21, v1, v21
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v29
-; GCN-NEXT: v_or_b32_e32 v23, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v3, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v6, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v10, v9, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v12, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 24, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v17, v12, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v12, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v12, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 24, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v25, v12, v8
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v12, v8
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v14, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v14, v12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v20, v14
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v24, v20
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v27, v24
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v27
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v29, v27
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v38, v19, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v19, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v19, v11
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v49, v19, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v19, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v19, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v23
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; kill: killed $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; kill: killed $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; kill: killed $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; kill: killed $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; kill: killed $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; kill: killed $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; kill: killed $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: .LBB44_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v29
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v21, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v19
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v6, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v13, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v1
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v56
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v5, v5, v8
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v1
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v9, v52, v9
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v1
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_or_b32_e32 v11, v22, v11
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v1
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v1, v13
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: v_mov_b32_e32 v2, v15
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v1
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v1, v15
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v1
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v1, v17
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v1
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v1, v19
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v1
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v1, v21
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v1
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v28, v25, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v37, v25, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v40, v2, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v45, v25, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v57, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v58, v25, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v59, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v60, v25, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v2, v24
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v2, v25
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v2, v26
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v2, v27
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v2, v29
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v2, v30
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v2, v31
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v2, v33
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v2, v34
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v2, v35
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v2, v36
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v2, v38
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v2, v50
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v2, v54
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v41, v2, v41
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v2, v42
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v2, v43
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v2, v44
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v2, v46
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v2, v47
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v56, 0xff, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v56, v2, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x300, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v61, v61, v2
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v62, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v63, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v1, v4
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v14
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v17
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s7, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v28
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s7, v32
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s7, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s7, v40
-; GCN-NEXT: v_add_i32_e32 v40, vcc, s7, v45
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s7, v57
-; GCN-NEXT: v_add_i32_e32 v57, vcc, s7, v58
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v60
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v40
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: v_or_b32_e32 v6, v48, v6
-; GCN-NEXT: v_or_b32_e32 v5, v49, v5
-; GCN-NEXT: v_or_b32_e32 v7, v51, v7
-; GCN-NEXT: v_or_b32_e32 v8, v52, v8
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: v_or_b32_e32 v10, v55, v10
-; GCN-NEXT: v_or_b32_e32 v11, v24, v11
-; GCN-NEXT: v_or_b32_e32 v12, v25, v12
-; GCN-NEXT: v_or_b32_e32 v13, v26, v13
-; GCN-NEXT: v_or_b32_e32 v14, v27, v14
-; GCN-NEXT: v_or_b32_e32 v15, v29, v15
-; GCN-NEXT: v_or_b32_e32 v16, v30, v16
-; GCN-NEXT: v_or_b32_e32 v17, v31, v17
-; GCN-NEXT: v_or_b32_e32 v18, v33, v18
-; GCN-NEXT: v_or_b32_e32 v19, v34, v19
-; GCN-NEXT: v_or_b32_e32 v20, v35, v20
-; GCN-NEXT: v_or_b32_e32 v21, v36, v21
-; GCN-NEXT: v_or_b32_e32 v22, v38, v22
-; GCN-NEXT: v_or_b32_e32 v23, v50, v23
-; GCN-NEXT: v_or_b32_e32 v24, v54, v28
-; GCN-NEXT: v_or_b32_e32 v25, v41, v32
-; GCN-NEXT: v_or_b32_e32 v26, v42, v37
-; GCN-NEXT: v_or_b32_e32 v27, v43, v40
-; GCN-NEXT: v_or_b32_e32 v28, v44, v45
-; GCN-NEXT: v_or_b32_e32 v29, v46, v57
-; GCN-NEXT: v_or_b32_e32 v30, v47, v58
-; GCN-NEXT: v_or_b32_e32 v31, v56, v59
-; GCN-NEXT: v_add_i32_e32 v35, vcc, s6, v61
-; GCN-NEXT: v_add_i32_e32 v49, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v38, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v30
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v31
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v22
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v20
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v16
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v15
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v14
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v14
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v13
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v7
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v6
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v9
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v12
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v11
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v8
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v37
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v51
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v5
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v48
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v39
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v50
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v35
-; GCN-NEXT: .LBB44_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_alignbit_b32 v11, v11, v13, 16
-; GCN-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v11, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v63, v5, v11, 16
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v56, v5, v11, 16
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v62, v5, v11, 16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 12, v0
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v61, v5, v11, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 20, v0
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v60, v13, v16, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_alignbit_b32 v16, v16, v19, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_alignbit_b32 v59, v22, v23, 16
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_alignbit_b32 v23, v23, v27, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 36, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_alignbit_b32 v58, v29, v35, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 40, v0
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_alignbit_b32 v3, v7, v3, 16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 44, v0
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_alignbit_b32 v57, v32, v35, 16
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 48, v0
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_alignbit_b32 v33, v34, v33, 16
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 52, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v35, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 56, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v6, 16
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 60, v0
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v52, 16
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 64, v0
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 0x44, v0
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_alignbit_b32 v18, v18, v54, 16
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x48, v0
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_alignbit_b32 v17, v36, v17, 16
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x4c, v0
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_alignbit_b32 v26, v26, v42, 16
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x50, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v25, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x54, v0
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_alignbit_b32 v37, v37, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x58, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v12, v14, v12, 16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x5c, v0
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v20, v20, v41, 16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x60, v0
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_alignbit_b32 v24, v28, v24, 16
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x64, v0
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_alignbit_b32 v30, v30, v40, 16
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x68, v0
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_alignbit_b32 v38, v39, v38, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x6c, v0
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_alignbit_b32 v48, v48, v55, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_alignbit_b32 v49, v50, v49, 16
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x74, v0
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_alignbit_b32 v51, v51, v53, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v63, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v62, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v60, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v59, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v51, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v23
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v27
-; VI-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v38
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v48
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v50
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB44_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: .LBB44_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB44_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v18, 0x300
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v0, v2, v0
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v2, v3, v2
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v5, v4
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v6, v5
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v6, v7, v6
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v35
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v36
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 3, v62
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v8, v9, v8
-; VI-NEXT: v_add_u16_e32 v9, 3, v63
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v59
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: v_add_u16_e32 v10, 3, v60
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v10, v11, v10
-; VI-NEXT: v_add_u16_e32 v11, 3, v58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v11, v12, v11
-; VI-NEXT: v_add_u16_e32 v12, 3, v47
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v46
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v12, v13, v12
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v13, v14, v13
-; VI-NEXT: v_add_u16_e32 v14, 3, v43
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v15, 3, v42
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: v_or_b32_e32 v14, v15, v14
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_e32 v16, v19, v16
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v40
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v55
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v21, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v48, 3, v48
-; VI-NEXT: v_or_b32_sdwa v48, v49, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v49, 3, v49
-; VI-NEXT: v_or_b32_sdwa v49, v50, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v20, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v50, 3, v50
-; VI-NEXT: v_or_b32_sdwa v50, v51, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v39
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v51, 3, v51
-; VI-NEXT: v_or_b32_sdwa v51, v52, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v52, 3, v52
-; VI-NEXT: v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v53, 3, v53
-; VI-NEXT: v_or_b32_sdwa v53, v54, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v54, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v19, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v39, v18
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v52
-; VI-NEXT: v_or_b32_e32 v19, v39, v19
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v50
-; VI-NEXT: v_or_b32_e32 v20, v39, v20
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v48
-; VI-NEXT: v_or_b32_e32 v21, v39, v21
-; VI-NEXT: v_or_b32_e32 v31, v31, v54
-; VI-NEXT: .LBB44_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v25
-; GFX9-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v39
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v49
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v52
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v3
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB44_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v17, v16, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v17, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v18, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v19, v20, v19, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v20, v21, v20, s6
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v28, v29, v28, s6
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: .LBB44_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB44_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_perm_b32 v0, v2, v0, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_perm_b32 v1, v1, v4, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v38, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_e32 v35, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v22
-; GFX9-NEXT: v_add_u16_e32 v36, 0x300, v36
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v33, 0x300, v21
-; GFX9-NEXT: v_add_u16_e32 v34, 0x300, v23
-; GFX9-NEXT: v_perm_b32 v29, v34, v29, s6
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v39
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v43
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v16
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v18
-; GFX9-NEXT: v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v20
-; GFX9-NEXT: v_perm_b32 v30, v33, v30, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v39, 0x300, v50
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v51
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v32, 0x300, v19
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v53
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v55
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v50, 0x300, v40
-; GFX9-NEXT: v_perm_b32 v21, v50, v21, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v41
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v51, 0x300, v42
-; GFX9-NEXT: v_perm_b32 v20, v51, v20, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v45
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v24
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v26
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v37
-; GFX9-NEXT: v_add_u16_e32 v37, 0x300, v38
-; GFX9-NEXT: v_add_u16_e32 v38, 0x300, v48
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v49
-; GFX9-NEXT: v_add_u16_e32 v48, 0x300, v52
-; GFX9-NEXT: v_add_u16_e32 v49, 0x300, v54
-; GFX9-NEXT: v_add_u16_e32 v52, 0x300, v44
-; GFX9-NEXT: v_add_u16_e32 v53, 0x300, v46
-; GFX9-NEXT: v_perm_b32 v18, v53, v18, s6
-; GFX9-NEXT: v_perm_b32 v19, v52, v19, s6
-; GFX9-NEXT: v_perm_b32 v22, v49, v22, s6
-; GFX9-NEXT: v_perm_b32 v23, v48, v23, s6
-; GFX9-NEXT: v_perm_b32 v24, v39, v24, s6
-; GFX9-NEXT: v_perm_b32 v25, v38, v25, s6
-; GFX9-NEXT: v_perm_b32 v26, v37, v26, s6
-; GFX9-NEXT: v_perm_b32 v27, v36, v27, s6
-; GFX9-NEXT: v_perm_b32 v28, v35, v28, s6
-; GFX9-NEXT: .LBB44_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:468
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:456
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v37, v30 :: v_dual_mov_b32 v54, v24
-; GFX11-NEXT: v_dual_mov_b32 v55, v28 :: v_dual_mov_b32 v52, v26
-; GFX11-NEXT: v_dual_mov_b32 v50, v22 :: v_dual_mov_b32 v53, v20
-; GFX11-NEXT: v_dual_mov_b32 v34, v18 :: v_dual_mov_b32 v51, v16
-; GFX11-NEXT: v_dual_mov_b32 v36, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v39, v4
-; GFX11-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v33, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v94, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v88, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v93, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v91, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v150, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v182, off, s32
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v45, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v46, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v47, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v61, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v62, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v63, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v72, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v73, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v79, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v89, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v90, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v95, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v104, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v105, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v132, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v161, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v160, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v176, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v167, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(62)
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v0
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v14
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v26
-; GFX11-NEXT: s_waitcnt vmcnt(61)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v88
-; GFX11-NEXT: s_waitcnt vmcnt(59)
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v93
-; GFX11-NEXT: s_waitcnt vmcnt(57)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v91
-; GFX11-NEXT: s_waitcnt vmcnt(55)
-; GFX11-NEXT: v_lshlrev_b16 v92, 8, v92
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v150
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v150, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v41, 8, v40
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v40, 8, v43
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v43, 8, v44
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v45
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v46
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v47
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v56
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v58
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v58, 8, v59
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v44, 8, v60
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v60, 8, v61
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v59, 8, v62
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v63
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v72
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v73
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v74
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v74, 8, v75
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v76
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v75, 8, v77
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v78
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v78, 8, v79
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v77, 8, v89
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v90
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v76, 8, v95
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v104
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v105
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v94
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v28
-; GFX11-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB44_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v48
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v70
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v71
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v84
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v82
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v51
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v34
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v66
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v128
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v113
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v132
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v100
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v161
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v160
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v176
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v167
-; GFX11-NEXT: v_perm_b32 v2, v2, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v5, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v7, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v9, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v11, v10, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v102
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v114
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v117
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v130
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v181
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v150
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v41
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v40
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v43
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v182
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v46
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v45
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v57
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v56
-; GFX11-NEXT: v_perm_b32 v7, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v12, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v14, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v16, v15, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v147
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v119
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v144
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v162
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v146
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v164
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v148
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v58
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v44
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v60
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v59
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v62
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v47
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v72
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v63
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v74
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v73
-; GFX11-NEXT: v_perm_b32 v12, v13, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v15, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v17, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v19, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v21, v20, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v166
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v145
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v163
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v179
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v165
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v183
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v180
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v65
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v75
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v61
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v78
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v77
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v79
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v76
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v90
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v89
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v92
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v91
-; GFX11-NEXT: v_perm_b32 v17, v18, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v20, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v22, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v24, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v26, v25, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v101
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v93
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v88
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v104
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v95
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v105
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v94
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v108
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v107
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v110
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v109
-; GFX11-NEXT: v_perm_b32 v22, v23, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v25, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v27, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v29, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v31, v30, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v103
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v99
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v129
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v131
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v116
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v111
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v106
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v122
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v121
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v123
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v120
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v125
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v124
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v126
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v127
-; GFX11-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v30, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v32, v31, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v34, v33, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v36, v35, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: .LBB44_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB44_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v118, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v131, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v116, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v129, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v35, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v126, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v127, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v125, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v124, v3
-; GFX11-NEXT: v_add_nc_u16 v33, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v116, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v98, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v1, v112, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_add_nc_u16 v3, v99, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v103, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v120, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v122, v1
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v121, v3
-; GFX11-NEXT: v_add_nc_u16 v99, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v111, v4
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v81, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v101, 3
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v86, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v97, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v83, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v106, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v110, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v109, v0
-; GFX11-NEXT: v_add_nc_u16 v83, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v108, v3
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v107, v4
-; GFX11-NEXT: v_add_nc_u16 v86, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v80, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v69, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v105, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v94, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v104, v2
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v95, v3
-; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v93, v4
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(26)
-; GFX11-NEXT: v_add_nc_u16 v0, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v65, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(24)
-; GFX11-NEXT: v_add_nc_u16 v3, v183, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v180, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v88, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v92, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v91, v1
-; GFX11-NEXT: v_add_nc_u16 v65, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v90, v3
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v89, v4
-; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_add_nc_u16 v1, v179, 3
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v165, 3
-; GFX11-NEXT: v_add_nc_u16 v80, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_add_nc_u16 v0, v177, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v163, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_add_nc_u16 v4, v166, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v79, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v76, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v78, v0
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v77, v3
-; GFX11-NEXT: v_add_nc_u16 v85, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v75, v4
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v145, 3
-; GFX11-NEXT: v_add_nc_u16 v97, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_add_nc_u16 v1, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v148, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_add_nc_u16 v3, v178, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v164, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v61, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v74, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v73, v2
-; GFX11-NEXT: v_add_nc_u16 v101, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v72, v3
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v63, v4
-; GFX11-NEXT: v_add_nc_u16 v103, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_add_nc_u16 v2, v162, 3
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v146, 3
-; GFX11-NEXT: v_add_nc_u16 v112, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(10)
-; GFX11-NEXT: v_add_nc_u16 v1, v149, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v144, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_add_nc_u16 v4, v147, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v62, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v47, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v60, v1
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v59, v3
-; GFX11-NEXT: v_add_nc_u16 v118, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v58, v4
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v119, 3
-; GFX11-NEXT: v_add_nc_u16 v119, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_add_nc_u16 v2, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v130, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_add_nc_u16 v3, v133, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v117, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v44, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v57, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v56, v0
-; GFX11-NEXT: v_add_nc_u16 v117, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v46, v3
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v45, v4
-; GFX11-NEXT: v_add_nc_u16 v129, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_add_nc_u16 v0, v114, 3
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v96, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v2, v102, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v87, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v55, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v43, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v182, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v41, v2
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v40, v3
-; GFX11-NEXT: v_add_nc_u16 v55, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v181, v4
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v37, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v52, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v53, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v50, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v150, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v176, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v167, v1
-; GFX11-NEXT: v_add_nc_u16 v50, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v161, v3
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v160, v4
-; GFX11-NEXT: v_add_nc_u16 v52, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v49, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v36, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v132, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v128, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v115, v0
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v100, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v113, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v39, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v32, 3
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_or_b32_e32 v35, v71, v35
-; GFX11-NEXT: v_or_b32_e32 v33, v82, v33
-; GFX11-NEXT: v_or_b32_e32 v0, v84, v0
-; GFX11-NEXT: v_or_b32_e32 v36, v70, v36
-; GFX11-NEXT: v_or_b32_e32 v32, v66, v32
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v38, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v36
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v39, 0x300, v1
-; GFX11-NEXT: v_perm_b32 v1, v33, v38, 0x5040100
-; GFX11-NEXT: v_perm_b32 v0, v35, v0, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v32, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v36, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v39, v34, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v52, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v50, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v96, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v129, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v12, v117, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v119, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v118, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v112, v15, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v103, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v17, v101, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v97, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v85, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v80, v20, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v69, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v22, v65, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v64, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v68, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v67, v25, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v86, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v27, v83, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v81, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v99, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v98, v30, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v116, v31, 0x5040100
-; GFX11-NEXT: .LBB44_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:516
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:580
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v41
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v40
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v46
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v48
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v39
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v35
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v33
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v1
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v42
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v44
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:128
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v45
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v12
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v13
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v14
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB45_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v28, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v24, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v25, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v26, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v30, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v35
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v40, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v39
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v42, v2, v3, 16
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v48
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v44, v2, v3, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v60
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v46, v2, v3, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v20
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v56, v2, v3, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v22
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v58, v2, v3, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v32
-; GCN-NEXT: v_alignbit_b32 v61, v2, v18, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v63
-; GCN-NEXT: v_alignbit_b32 v15, v2, v19, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v21
-; GCN-NEXT: v_alignbit_b32 v14, v2, v17, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v31
-; GCN-NEXT: v_alignbit_b32 v13, v2, v27, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v29
-; GCN-NEXT: v_alignbit_b32 v59, v2, v33, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v35
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v39
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v48
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v60
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v20
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v22
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v34
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v9
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v11
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v43
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v47
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 24, v62
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 24, v16
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v23
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v51, v52, v16, 16
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v35, v12, v16, 16
-; GCN-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v36, v36, v12, 16
-; GCN-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v37, v37, v12, 16
-; GCN-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v38, v38, v12, 16
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v39, v6, v12, 16
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v48, v3, v12, 16
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 16
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v50, v5, v3, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v52
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v52, v8, v3, 16
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v53, v10, v3, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v54, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v55, v7, v2, 16
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v41, v9, v2, 16
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v45, v11, v2, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v57, v1, v2, 16
-; GCN-NEXT: v_alignbit_b32 v4, v51, v28, 24
-; GCN-NEXT: v_alignbit_b32 v10, v51, v28, 16
-; GCN-NEXT: v_alignbit_b32 v3, v51, v28, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v24, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v24, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v8, v35, v24, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v36, v25, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v36, v25, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v12, v36, v25, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v52
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: .LBB45_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB45_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v29
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_alignbit_b32 v59, v13, v12, 16
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v31
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v21
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_alignbit_b32 v14, v15, v14, 16
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v63
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_alignbit_b32 v15, v17, v15, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v16
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v18
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v47
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v43
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v34
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v7
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v7
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v20
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v42, 0xffff0000, v7
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v60
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v62, 0xffff0000, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v7
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v4
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v60, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v18
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v19
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v23
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v27
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v31
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v33
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v53
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v54
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v55
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v40
-; GCN-NEXT: v_add_f32_e32 v50, 0x40c00000, v41
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v42
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v43
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v44
-; GCN-NEXT: v_add_f32_e32 v52, 0x40c00000, v45
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v46
-; GCN-NEXT: v_add_f32_e32 v39, 0x40c00000, v47
-; GCN-NEXT: v_add_f32_e32 v51, 0x40c00000, v56
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v57
-; GCN-NEXT: v_add_f32_e32 v47, 0x40c00000, v58
-; GCN-NEXT: v_add_f32_e32 v48, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v53, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v61
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v62
-; GCN-NEXT: v_add_f32_e32 v49, 0x40c00000, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v61, v1, v6, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GCN-NEXT: v_alignbit_b32 v58, v1, v2, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v24
-; GCN-NEXT: v_alignbit_b32 v56, v1, v3, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v29
-; GCN-NEXT: v_alignbit_b32 v46, v1, v4, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v31
-; GCN-NEXT: v_alignbit_b32 v44, v1, v5, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v7
-; GCN-NEXT: v_alignbit_b32 v42, v1, v33, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v28
-; GCN-NEXT: v_alignbit_b32 v40, v1, v8, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v26
-; GCN-NEXT: v_alignbit_b32 v30, v1, v30, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v50
-; GCN-NEXT: v_alignbit_b32 v26, v1, v25, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v52
-; GCN-NEXT: v_alignbit_b32 v25, v1, v41, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v54
-; GCN-NEXT: v_alignbit_b32 v24, v1, v51, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v55
-; GCN-NEXT: v_alignbit_b32 v28, v1, v53, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 24, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 24, v20
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 24, v60
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 24, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v57, v20, v21, 16
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v45, v18, v20, 16
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v41, v10, v18, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v55, v1, v10, 16
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v54, v62, v1, 16
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v53, v12, v1, 16
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v52, v11, v1, 16
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_alignbit_b32 v50, v4, v1, 16
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v3, v17, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v48, v2, v19, 16
-; GCN-NEXT: v_alignbit_b32 v39, v6, v16, 16
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v38, v29, v22, 16
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v37, v8, v23, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v36, v7, v43, 16
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v35, v5, v47, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v31
-; GCN-NEXT: v_alignbit_b32 v51, v5, v9, 16
-; GCN-NEXT: v_alignbit_b32 v4, v51, v28, 24
-; GCN-NEXT: v_alignbit_b32 v10, v51, v28, 16
-; GCN-NEXT: v_alignbit_b32 v3, v51, v28, 8
-; GCN-NEXT: v_alignbit_b32 v1, v35, v24, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v24, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v8, v35, v24, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v36, v25, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v36, v25, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v12, v36, v25, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v26, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v38, v30, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v40, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v48, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v44, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v46, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v52, v56, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v61, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v15, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v14, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v13, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v59, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v52
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB45_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v51
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v4
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v8
-; GCN-NEXT: v_or_b32_e32 v29, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v12
-; GCN-NEXT: v_or_b32_e32 v2, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v36
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v62, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v30
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v40
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v39
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v42
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v48
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v44
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v49
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v46
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v50
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v18, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v56
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v52
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v58
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v53
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v22, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v61
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v15
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v55
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v25, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v14, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v41
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v26, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v13
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v45
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v27, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v57
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v28, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v30, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v49, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v4, v1, v30
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v5, v1, v33
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v62
-; GCN-NEXT: v_or_b32_e32 v61, v3, v34
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v36
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v29, v29, v37
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v30, v30, v38
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v48
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v9, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v11, v11, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v17, v17, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v19, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v21, v21, v42
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v43
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v23, v23, v44
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v45
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v46
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v47
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v56
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v13, v57
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v58
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v12, v59
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v60
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; kill: killed $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB45_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v29
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v28
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v28
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v28
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v27
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v27
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v26
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v26
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v26
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v5
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v4
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v3
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v2
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v17
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v1
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v17
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[31:32]
-; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v12
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v45, v46
-; VI-NEXT: v_lshrrev_b64 v[46:47], 24, v[29:30]
-; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v31
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v46, v63
-; VI-NEXT: v_mov_b32_e32 v63, v50
-; VI-NEXT: v_lshrrev_b64 v[50:51], 24, v[27:28]
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v57, 16, v9
-; VI-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v8
-; VI-NEXT: v_mov_b32_e32 v51, v57
-; VI-NEXT: v_mov_b32_e32 v50, v56
-; VI-NEXT: v_lshrrev_b64 v[56:57], 24, v[25:26]
-; VI-NEXT: v_mov_b32_e32 v57, v43
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[17:18]
-; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v40, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v60, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v54, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v32
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v31
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v30
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v30
-; VI-NEXT: v_lshrrev_b32_e32 v53, 8, v30
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: v_mov_b32_e32 v47, v34
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: .LBB45_2: ; %Flow
-; VI-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: s_xor_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB45_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v18
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, 0x7fff, v34
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v18, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v18
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v18, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v17
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_cndmask_b32_e32 v17, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v17, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v20
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v20
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v20, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v19
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v19
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v19, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v22
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v22
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v22, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v21
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v21
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v21, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v24
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v24
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v24, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v23
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v23
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v23, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v26
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v26
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v26, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v25
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v25
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v25, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v28
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v28
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v28, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v27
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v27
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v27, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v30
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v30
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v30, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v29
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v29
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v29, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v32
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v32
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; VI-NEXT: v_alignbit_b32 v32, v32, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v31
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v31
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v31, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v2
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v2
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v1
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v1
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v4
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v3
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v3
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v6
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v6
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v5
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v5
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v8
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v8
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v7
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v7
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v10
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v10
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v9
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v9
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v12
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v12
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v11
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v11
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v14
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v13
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v13
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v16
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v16
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: v_alignbit_b32 v16, v16, v33, 16
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v15
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v15, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v15
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v34, v35, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v33, 16
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[3:4]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[1:2]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[17:18]
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v28
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v28
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v28
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v27
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v27
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v26
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v26
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v26
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v25
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v24
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v23
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v22
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v5
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v21
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v4
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v3
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v19
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v18
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v2
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v18
-; VI-NEXT: v_lshrrev_b64 v[56:57], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v45, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v40, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v50, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v35, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v60, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v54, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v6
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v32
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v48, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v31
-; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v31
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v30
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v30
-; VI-NEXT: v_lshrrev_b32_e32 v53, 8, v30
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v29
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v17
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: .LBB45_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v44, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v1, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v44, 8, v43
-; VI-NEXT: v_or_b32_sdwa v2, v2, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v43
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v43, v44, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v43, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v38
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v54
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v59, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v41
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v50
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v40
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v46
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v45
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v56
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v53
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v63
-; VI-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v48
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v46, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: v_mov_b32_e32 v47, v16
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; kill: killed $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; kill: killed $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; kill: killed $vgpr59
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(34)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; kill: killed $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB45_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v2
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(36)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v47
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(38)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v47
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v47
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v46
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v46
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 24, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v18
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v8
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[46:47]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v20
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[7:8]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[3:4]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[1:2]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[58:59], 24, v[17:18]
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v24
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v23
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v23
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v22
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v22
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v21
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v21
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v17
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB45_2: ; %Flow
-; GFX9-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v58, v57
-; GFX9-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB45_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v18
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v33, v16, v33, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v18, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v34, v15, v33, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v17
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v33, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v20
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v18, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v19
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v17, v15, v20, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v22
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v18, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v21
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v61, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v17, v15, v61, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v24
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v59, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v23
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v58, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v26
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_perm_b32 v63, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v25
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v62, v15, v18, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v28
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v60, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v27
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v56, v15, v26, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v30
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v29
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v33, v15, v25, s7
-; GFX9-NEXT: s_waitcnt vmcnt(52)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v32
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: s_waitcnt vmcnt(51)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v31
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_perm_b32 v35, v15, v24, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_bfe_u32 v15, v2, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v16, v17, vcc
-; GFX9-NEXT: v_add3_u32 v15, v15, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v2, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_add3_u32 v15, v15, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v15, v16, vcc
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v15, vcc
-; GFX9-NEXT: v_perm_b32 v37, v1, v23, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v2, v15, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v2, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v2, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_perm_b32 v48, v1, v20, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_perm_b32 v50, v1, v17, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v7, vcc
-; GFX9-NEXT: v_perm_b32 v52, v1, v4, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v10
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v2, v7, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v2, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v9
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v2, v15, vcc
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v15, vcc
-; GFX9-NEXT: v_perm_b32 v39, v1, v3, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v12
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v12, v1, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v9, v15, vcc
-; GFX9-NEXT: v_add3_u32 v12, v12, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v12, v15, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v1, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v15, v15, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v15, v16, vcc
-; GFX9-NEXT: v_perm_b32 v54, v11, v2, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_bfe_u32 v15, v11, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v15, v15, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v41, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_bfe_u32 v15, v13, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v16, v41, vcc
-; GFX9-NEXT: v_add3_u32 v15, v15, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v15, v16, vcc
-; GFX9-NEXT: v_perm_b32 v41, v13, v1, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v47
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_bfe_u32 v15, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v15, v16, vcc
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v47
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v44, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v44, v16, v44, vcc
-; GFX9-NEXT: v_perm_b32 v16, v44, v13, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v14
-; GFX9-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v10
-; GFX9-NEXT: v_perm_b32 v53, v8, v5, s7
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v8, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v32
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v30
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v28
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v34, v30, v27, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v27
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GFX9-NEXT: v_perm_b32 v36, v32, v29, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v29
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v25
-; GFX9-NEXT: v_perm_b32 v38, v22, v31, s7
-; GFX9-NEXT: v_perm_b32 v42, v14, v11, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v31
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v19
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v22
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GFX9-NEXT: v_perm_b32 v55, v12, v9, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v20
-; GFX9-NEXT: v_perm_b32 v49, v19, v21, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v46
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v45, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v45, v45, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v43, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v46
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v43, v45, v43, vcc
-; GFX9-NEXT: v_bfe_u32 v45, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v45, v45, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v46, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v45, v46, vcc
-; GFX9-NEXT: v_perm_b32 v15, v15, v43, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v44
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v43
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[15:16]
-; GFX9-NEXT: v_perm_b32 v51, v6, v18, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v18
-; GFX9-NEXT: v_perm_b32 v40, v10, v7, s7
-; GFX9-NEXT: v_perm_b32 v57, v28, v60, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v21
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v60
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v61
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v39
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v17
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[41:42]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[54:55]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[39:40]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[52:53]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[50:51]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[48:49]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[37:38]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[35:36]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[33:34]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[56:57]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[62:63]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[58:59]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[60:61]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[43:44]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b64 v[43:44], 24, v[43:44]
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v16
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v42
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v42
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v41
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v41
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v55
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v55
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v54
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v40
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v39
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v53
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v53
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 24, v51
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 8, v51
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 24, v38
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 8, v38
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v37
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v37
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v36
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v36
-; GFX9-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v35
-; GFX9-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v34
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v33
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v57
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v57
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v56
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v56
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v63
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v63
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v62
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v50
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v62
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v50
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v49
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v59
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v59
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v61
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v61
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v60
-; GFX9-NEXT: v_mov_b32_e32 v33, v60
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v54
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 8, v40
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v52
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v49
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v48
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v48
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v58
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 8, v58
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v33
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v62
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v62
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v61
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v61
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 24, v60
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v60
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v59
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v59
-; GFX9-NEXT: .LBB45_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v40
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v51
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v48
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v42
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v52
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v60
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v41
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v47, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v51
-; GFX9-NEXT: v_or_b32_sdwa v38, v38, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v53
-; GFX9-NEXT: v_or_b32_sdwa v15, v46, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v38, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v34, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v56, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v50
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v58
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v63
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x15
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:96
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:92
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB45_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v76, 8, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v17
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b64 v[67:68], 24, v[17:18]
-; GFX11-NEXT: .LBB45_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB45_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-NEXT: v_add_f32_e32 v35, 0x40c00000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v33, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v33
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_add3_u32 v17, v17, v33, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v17, v17, v36 :: v_dual_and_b32 v18, 0xffff0000, v18
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v37, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_add3_u32 v37, v37, v18, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v77, v37, v39 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v20
-; GFX11-NEXT: v_perm_b32 v69, v77, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v38, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v34
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 24, v69
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v38, v38, v34, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v69
-; GFX11-NEXT: v_dual_cndmask_b32 v33, v38, v18 :: v_dual_add_f32 v20, 0x40c00000, v20
-; GFX11-NEXT: v_bfe_u32 v48, v35, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v35
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v37
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v36, v48, v35, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v35, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v36, v39, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v19
-; GFX11-NEXT: v_add3_u32 v35, v35, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_perm_b32 v68, v34, v33, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v34, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v35, v37, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v38, 0x40c00000, v19 :: v_dual_lshlrev_b32 v37, 16, v22
-; GFX11-NEXT: v_add3_u32 v34, v34, v20, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-NEXT: v_bfe_u32 v35, v38, 16, 1
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v68
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v68
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v34, v19, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v39, v36, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v36
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_add3_u32 v35, v35, v38, 0x7fff
-; GFX11-NEXT: v_perm_b32 v65, v19, v18, 0x7060302
-; GFX11-NEXT: v_add3_u32 v20, v39, v36, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v19
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 24, v65
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v20, v34, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v37
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 16, v34
-; GFX11-NEXT: v_bfe_u32 v36, v20, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v35, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-NEXT: v_add3_u32 v36, v36, v20, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_perm_b32 v64, v35, v34, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 8, v65
-; GFX11-NEXT: v_bfe_u32 v48, v37, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v36, v38, vcc_lo
-; GFX11-NEXT: v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_add_f32 v39, 0x40c00000, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v64
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v64
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[64:65]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v35, v22, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v22
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v24
-; GFX11-NEXT: v_bfe_u32 v36, v39, 16, 1
-; GFX11-NEXT: v_add3_u32 v35, v35, v22, 0x7fff
-; GFX11-NEXT: v_add3_u32 v22, v48, v37, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v39
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_add3_u32 v36, v36, v39, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v35, v21, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v37
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v71, v21, v20, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v21
-; GFX11-NEXT: v_dual_cndmask_b32 v35, v22, v35 :: v_dual_add_f32 v22, 0x40c00000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 16, v35
-; GFX11-NEXT: v_bfe_u32 v37, v22, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v36, v48, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v22
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-NEXT: v_add3_u32 v37, v37, v22, 0x7fff
-; GFX11-NEXT: v_perm_b32 v70, v36, v35, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v48, 0x40c00000, v23
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v49, v38, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v22, v37, v39 :: v_dual_lshlrev_b32 v39, 16, v26
-; GFX11-NEXT: v_bfe_u32 v36, v24, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v24
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_bfe_u32 v37, v48, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_add3_u32 v36, v36, v24, 0x7fff
-; GFX11-NEXT: v_add3_u32 v24, v49, v38, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v48
-; GFX11-NEXT: v_add3_u32 v37, v37, v48, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v26, 0x40c00000, v26 :: v_dual_cndmask_b32 v23, v36, v23
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 16, v70
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v23
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v24, v36, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v39
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_perm_b32 v81, v23, v22, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v24, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v37, v37, v49, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v24
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_add3_u32 v38, v38, v24, 0x7fff
-; GFX11-NEXT: v_perm_b32 v80, v37, v36, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v37, v26, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v49, 0x40c00000, v25
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v26
-; GFX11-NEXT: v_bfe_u32 v50, v39, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v38, v48, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v37, v26, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: v_bfe_u32 v38, v49, 16, 1
-; GFX11-NEXT: v_add3_u32 v26, v50, v39, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v28
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v49
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v37, v25, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v39
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_add3_u32 v38, v38, v49, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 16, v80
-; GFX11-NEXT: v_dual_cndmask_b32 v37, v26, v37 :: v_dual_add_f32 v26, 0x40c00000, v48
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_bfe_u32 v39, v26, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v38, v38, v50, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v26
-; GFX11-NEXT: v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: v_add3_u32 v39, v39, v26, 0x7fff
-; GFX11-NEXT: v_perm_b32 v82, v38, v37, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v38, v28, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v27
-; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v28
-; GFX11-NEXT: v_bfe_u32 v51, v48, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v26, v39, v49, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v38, v28, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: v_bfe_u32 v39, v50, 16, 1
-; GFX11-NEXT: v_add3_u32 v28, v51, v48, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v30
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v50
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v38, v27, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v48
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_add3_u32 v39, v39, v50, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_perm_b32 v83, v25, v24, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v38, v28, v38, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v49
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v49, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v48, v28, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v39, v39, v51, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v28
-; GFX11-NEXT: v_add_f32_e32 v49, 0x40c00000, v49
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: v_add3_u32 v48, v48, v28, 0x7fff
-; GFX11-NEXT: v_perm_b32 v84, v39, v38, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v39, v30, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v51, 0x40c00000, v29
-; GFX11-NEXT: v_or_b32_e32 v29, 0x400000, v30
-; GFX11-NEXT: v_bfe_u32 v52, v49, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v48, v50, vcc_lo
-; GFX11-NEXT: v_add3_u32 v39, v39, v30, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: v_bfe_u32 v48, v51, 16, 1
-; GFX11-NEXT: v_add3_u32 v30, v52, v49, 0x7fff
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v32
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v51
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v39, v29, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v49
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_add3_u32 v48, v48, v51, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 24, v83
-; GFX11-NEXT: v_dual_cndmask_b32 v39, v30, v39 :: v_dual_add_f32 v30, 0x40c00000, v50
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v31
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: v_bfe_u32 v49, v30, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v48, v48, v52, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v30
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v50
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: v_add3_u32 v49, v49, v30, 0x7fff
-; GFX11-NEXT: v_perm_b32 v86, v48, v39, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v48, v32, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v52, 0x40c00000, v31
-; GFX11-NEXT: v_or_b32_e32 v31, 0x400000, v32
-; GFX11-NEXT: v_bfe_u32 v53, v50, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v30, v49, v51, vcc_lo
-; GFX11-NEXT: v_add3_u32 v48, v48, v32, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_bfe_u32 v49, v52, 16, 1
-; GFX11-NEXT: v_add3_u32 v32, v53, v50, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v2
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v52
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v48, v31, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v50
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_add3_u32 v49, v49, v52, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_perm_b32 v85, v27, v26, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v48, v32, v48, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v32, 0x40c00000, v51
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v51, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v50, v32, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v49, v49, v53, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v32
-; GFX11-NEXT: v_add_f32_e32 v51, 0x40c00000, v51
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_add3_u32 v50, v50, v32, 0x7fff
-; GFX11-NEXT: v_perm_b32 v96, v49, v48, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v49, v2, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v53, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v55, v51, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v50, v52, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v49, v2, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_bfe_u32 v32, v53, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v51
-; GFX11-NEXT: v_perm_b32 v87, v29, v28, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v49, v54, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v55, v51, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: v_add3_u32 v52, v32, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v53
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v49, v50, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v49, 0x40c00000, v54
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v51, v49, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v50, v52, v55, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v49
-; GFX11-NEXT: v_add3_u32 v51, v51, v49, 0x7fff
-; GFX11-NEXT: v_perm_b32 v98, v50, v32, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v50, v4, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v52, 0x40c00000, v52
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v4
-; GFX11-NEXT: v_add3_u32 v50, v50, v4, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v55, v52, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v49, v51, v53, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v51, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v52
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 16, v98
-; GFX11-NEXT: v_lshrrev_b32_e32 v76, 8, v98
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v50, v54, vcc_lo
-; GFX11-NEXT: v_add3_u32 v50, v55, v52, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_add3_u32 v51, v51, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v52, 0x40c00000, v54
-; GFX11-NEXT: v_cndmask_b32_e32 v50, v50, v53, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v53, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v52
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v51, v55, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v51, v52, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v53, 0x40c00000, v53
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_perm_b32 v100, v3, v50, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v3, v6, 16, 1
-; GFX11-NEXT: v_add3_u32 v51, v51, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v6
-; GFX11-NEXT: v_bfe_u32 v66, v53, 16, 1
-; GFX11-NEXT: v_bfe_u32 v52, v5, 16, 1
-; GFX11-NEXT: v_add3_u32 v3, v3, v6, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v51, v51, v54, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v6, v66, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v53
-; GFX11-NEXT: v_add3_u32 v52, v52, v5, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v5
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v3, v55, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v97, v31, v30, 0x7060302
-; GFX11-NEXT: v_perm_b32 v103, v3, v51, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 16, v3
-; GFX11-NEXT: v_dual_cndmask_b32 v53, v6, v54 :: v_dual_add_f32 v6, 0x40c00000, v55
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v54, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v52, v66, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v52, v6, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v54, 0x40c00000, v54
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_perm_b32 v102, v5, v53, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v5, v8, 16, 1
-; GFX11-NEXT: v_add3_u32 v52, v52, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v8
-; GFX11-NEXT: v_bfe_u32 v67, v54, 16, 1
-; GFX11-NEXT: v_perm_b32 v99, v2, v1, 0x7060302
-; GFX11-NEXT: v_add3_u32 v5, v5, v8, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v52, v55, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_bfe_u32 v52, v7, 16, 1
-; GFX11-NEXT: v_add3_u32 v8, v67, v54, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v54
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v7
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v5, v66, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v66, 16, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v54, v54
-; GFX11-NEXT: v_add3_u32 v52, v52, v7, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-NEXT: v_cndmask_b32_e32 v54, v8, v55, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v66
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v55, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v52, v67, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v52, v8, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v55, 0x40c00000, v55
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_perm_b32 v182, v7, v54, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v7, v10, 16, 1
-; GFX11-NEXT: v_add3_u32 v52, v52, v8, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v10
-; GFX11-NEXT: v_bfe_u32 v112, v55, 16, 1
-; GFX11-NEXT: v_perm_b32 v183, v5, v6, 0x7060302
-; GFX11-NEXT: v_add3_u32 v7, v7, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v52, v66, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v52, v9, 16, 1
-; GFX11-NEXT: v_add3_u32 v10, v112, v55, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v55
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v9
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v7, v67, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55
-; GFX11-NEXT: v_add3_u32 v52, v52, v9, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: v_perm_b32 v101, v4, v49, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 16, v4
-; GFX11-NEXT: v_dual_cndmask_b32 v55, v10, v66 :: v_dual_add_f32 v10, 0x40c00000, v67
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v49
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v50
-; GFX11-NEXT: v_bfe_u32 v66, v10, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v52, v112 :: v_dual_lshlrev_b32 v52, 16, v11
-; GFX11-NEXT: v_bfe_u32 v67, v12, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v176, v9, v55, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v52
-; GFX11-NEXT: v_add3_u32 v52, v66, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v10
-; GFX11-NEXT: v_add3_u32 v67, v67, v12, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v12
-; GFX11-NEXT: v_bfe_u32 v113, v9, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v11, 0x40c00000, v11 :: v_dual_cndmask_b32 v10, v52, v66
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v9
-; GFX11-NEXT: v_add3_u32 v52, v113, v9, 0x7fff
-; GFX11-NEXT: v_perm_b32 v177, v7, v8, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v32
-; GFX11-NEXT: v_dual_cndmask_b32 v12, v67, v112 :: v_dual_lshlrev_b32 v67, 16, v14
-; GFX11-NEXT: v_bfe_u32 v112, v11, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v26
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v52, v66 :: v_dual_add_f32 v52, 0x40c00000, v67
-; GFX11-NEXT: v_add3_u32 v66, v112, v11, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v112, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_bfe_u32 v113, v52, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v24
-; GFX11-NEXT: v_dual_cndmask_b32 v11, v66, v67 :: v_dual_add_f32 v66, 0x40c00000, v112
-; GFX11-NEXT: v_add3_u32 v67, v113, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v52
-; GFX11-NEXT: v_bfe_u32 v113, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_bfe_u32 v114, v66, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v115, 0x400000, v66
-; GFX11-NEXT: v_perm_b32 v162, v11, v9, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v52, v67, v112, vcc_lo
-; GFX11-NEXT: v_add3_u32 v67, v113, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v14
-; GFX11-NEXT: v_add3_u32 v113, v114, v66, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v114, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_bfe_u32 v116, v13, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v67, v112 :: v_dual_add_f32 v67, 0x40c00000, v114
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v66, v66
-; GFX11-NEXT: v_add3_u32 v112, v116, v13, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v149, v14, v52, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v114, v67, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v66, v113, v115, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v113, 0x400000, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v115, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_perm_b32 v163, v12, v10, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_dual_cndmask_b32 v13, v112, v113 :: v_dual_add_f32 v112, 0x40c00000, v115
-; GFX11-NEXT: v_add3_u32 v113, v114, v67, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v114, 0x400000, v67
-; GFX11-NEXT: v_bfe_u32 v115, v16, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v67, v67
-; GFX11-NEXT: v_bfe_u32 v116, v112, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v117, 0x400000, v112
-; GFX11-NEXT: v_or_b32_e32 v118, 0x400000, v15
-; GFX11-NEXT: v_perm_b32 v148, v13, v66, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v67, v113, v114, vcc_lo
-; GFX11-NEXT: v_add3_u32 v114, v115, v16, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v115, 0x400000, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_bfe_u32 v113, v15, 16, 1
-; GFX11-NEXT: v_add3_u32 v116, v116, v112, 0x7fff
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 16, v33
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[96:97]
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v114, v115, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v112, v112
-; GFX11-NEXT: v_add3_u32 v113, v113, v15, 0x7fff
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[86:87]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[84:85]
-; GFX11-NEXT: v_perm_b32 v135, v16, v67, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v112, v116, v117, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v52
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v53
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v113, v118, vcc_lo
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v12, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v8, 16, v6
-; GFX11-NEXT: v_perm_b32 v134, v15, v112, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 16, v112
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 16, v48
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v51
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[134:135]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[148:149]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[162:163]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[176:177]
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v37
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[182:183]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[82:83]
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 16, v67
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 16, v66
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 16, v54
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[102:103]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[98:99]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[80:81]
-; GFX11-NEXT: v_lshrrev_b32_e32 v11, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 16, v55
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v39
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[100:101]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[70:71]
-; GFX11-NEXT: v_lshrrev_b64 v[67:68], 24, v[68:69]
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 24, v135
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 8, v135
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v134
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v134
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 24, v149
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v149
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v148
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 8, v148
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 24, v163
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 8, v163
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 16, v162
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 8, v162
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 24, v177
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v177
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 16, v176
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 8, v176
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 24, v183
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 8, v183
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v182
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v182
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 24, v103
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 8, v103
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 16, v102
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v102
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 24, v101
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 8, v101
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 16, v100
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 8, v100
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 24, v99
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 8, v99
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 24, v97
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 8, v97
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 16, v96
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 8, v96
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 24, v87
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v87
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v86
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v86
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 24, v85
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v85
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 16, v84
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v84
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v83
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 16, v82
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v82
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 24, v81
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v81
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v80
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 24, v71
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v71
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v70
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 16, v77
-; GFX11-NEXT: .LBB45_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v76
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v63
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v75
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v66, 0xff, v60
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v73
-; GFX11-NEXT: v_or_b32_e32 v54, v66, v54
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v56
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v166
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v65
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v62
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v65
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v160
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v51
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v54, v65, v66
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v45
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v44
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v55
-; GFX11-NEXT: v_or_b32_e32 v53, v65, v53
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v147
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v41
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v55, v65
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v181
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v54
-; GFX11-NEXT: v_or_b32_e32 v52, v55, v52
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v180
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v132
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v179
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v53
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v52, v54, v55
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v167
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v177
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v51, v54, v51
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v52
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v51, 0xff, v119
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v164
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v162
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v161
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v150
-; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v149
-; GFX11-NEXT: v_or_b32_e32 v51, v51, v52
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v53
-; GFX11-NEXT: v_or_b32_e32 v50, v54, v50
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v65, v66
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v51
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v50
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v52
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v145
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v144
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v116
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v135
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v134
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v131
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v114
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v74
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v72
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v48
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v61
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v69
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v59
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v47
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v46
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v146
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v43
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v70
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v40
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v71
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v182
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v178
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v81
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v128
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v176
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v163
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v83
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v117
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v151
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v103
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v115
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v102
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v86
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v101
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v100
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v99
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v97
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v112
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v98
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x15
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:92
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:96
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v50, v27
-; GCN-NEXT: v_mov_b32_e32 v49, v25
-; GCN-NEXT: v_mov_b32_e32 v39, v21
-; GCN-NEXT: v_mov_b32_e32 v48, v3
-; GCN-NEXT: v_mov_b32_e32 v37, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:68
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:64
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:56
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:48
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v4
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v10
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v12
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v14
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v20
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 8, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 8, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:392
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:112
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:124
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v21
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v34
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v27
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v8
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:136
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:144
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:156
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v4
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:160
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:172
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:168
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:176
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:188
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v4
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:192
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(4) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:240
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:252
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:272
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:284
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:296
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:304
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:316
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v3
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v4
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:348
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v3
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 8, v4
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:360
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:372
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:384
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v2
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:152
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:184
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:216
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:248
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:280
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:312
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:344
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:376
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:368
-; GCN-NEXT: s_waitcnt vmcnt(13)
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 8, v3
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB46_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_mov_b32_e32 v26, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v37
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v48
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v7
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v9
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v11
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v13
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v15
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v0, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_or_b32_e32 v0, v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v19
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v0, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v39
-; GCN-NEXT: v_or_b32_e32 v25, v0, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v0, v24
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v49
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v0, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v50
-; GCN-NEXT: v_or_b32_e32 v33, v0, v28
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v29
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v0, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v12
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v0, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v14
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v2, v4
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v2, v4
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v2, v4
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v2, v4
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v2, v4
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v4
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v6, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v6, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v51
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: v_mov_b32_e32 v7, v8
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v12, v5
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v30
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v12, v5
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v7
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v5
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v53, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v18, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v5
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v40, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v5
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v26
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v13
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v42, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v15
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v43, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v15, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v17, v27
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v46, v27
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v51, v51, v52
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v52, v52, v54
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v54, v54, v55
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v56
-; GCN-NEXT: v_or_b32_e32 v55, v55, v41
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v41, v41, v44
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v44, v44, v45
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v45, v45, v47
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v58
-; GCN-NEXT: v_or_b32_e32 v47, v47, v57
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v56, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v56, v56, v62
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v27
-; GCN-NEXT: v_or_b32_e32 v31, v57, v31
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v60
-; GCN-NEXT: v_or_b32_e32 v57, v57, v63
-; GCN-NEXT: v_and_b32_e32 v58, 0xff, v61
-; GCN-NEXT: v_or_b32_e32 v1, v58, v1
-; GCN-NEXT: v_and_b32_e32 v58, 0xff, v59
-; GCN-NEXT: v_or_b32_e32 v3, v58, v3
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v25
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v23
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v32
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v33
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v34
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v36
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v37
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v38
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v39
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v48
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v4
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v49
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v50
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v6
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v8
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v14
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v16
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v53
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v9
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v18
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v20
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v40
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v5
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v42
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v43
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v46
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v3
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: .LBB46_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB46_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v59
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v6, v3, v2
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v61
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v10, v1, v2
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v60
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v16, v63, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v20, v31, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v62, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v58
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v57, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v47, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v45, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v44, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v56
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v41, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v55, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v54, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v1, v52, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v46
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v3, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v0, v24
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v40
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v0, v26
-; GCN-NEXT: v_mov_b32_e32 v0, v37
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v4, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v53
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v8
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v1, v34
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v21, v39
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 3, v51
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v1, v48
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v48
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v25, v49
-; GCN-NEXT: v_mov_b32_e32 v27, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v49
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v50
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v50
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v51
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v51
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v52
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v52
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v53
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v54
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v54
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v55
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v2, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 3, v12
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v2, v41
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v29
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v42
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v42
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v4, v43
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v25
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v2, v44
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v45, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v45
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v45
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v21
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v46
-; GCN-NEXT: v_or_b32_e32 v46, v22, v46
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v19
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v2, v47
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 3, v17
-; GCN-NEXT: v_and_b32_e32 v56, 0xff, v56
-; GCN-NEXT: v_or_b32_e32 v56, v18, v56
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 3, v15
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v57, v2, v57
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 3, v13
-; GCN-NEXT: v_and_b32_e32 v58, 0xff, v58
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v58, v2, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 3, v11
-; GCN-NEXT: v_and_b32_e32 v59, 0xff, v59
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v59, v2, v59
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 3, v9
-; GCN-NEXT: v_and_b32_e32 v60, 0xff, v60
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v2, v60
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 3, v7
-; GCN-NEXT: v_and_b32_e32 v61, 0xff, v61
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v2, v61
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 3, v5
-; GCN-NEXT: v_and_b32_e32 v62, 0xff, v62
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v2, v62
-; GCN-NEXT: v_add_i32_e32 v63, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v63, 0xff, v63
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v63
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v45, v0, v3
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 0x300, v6
-; GCN-NEXT: v_add_i32_e32 v43, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v42, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s6, v20
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v40, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v55, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v53, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v49, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v38, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v35, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v24
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v26
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v8
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v12
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v46, vcc, s6, v46
-; GCN-NEXT: v_add_i32_e32 v47, vcc, s6, v47
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s6, v56
-; GCN-NEXT: v_add_i32_e32 v57, vcc, s6, v57
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s6, v58
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s6, v59
-; GCN-NEXT: v_add_i32_e32 v60, vcc, s6, v60
-; GCN-NEXT: v_add_i32_e32 v61, vcc, s6, v61
-; GCN-NEXT: v_add_i32_e32 v62, vcc, s6, v62
-; GCN-NEXT: v_add_i32_e32 v63, vcc, s6, v63
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s6, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v63
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v62
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v61
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v60
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v59
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v58
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v57
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v56
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v47
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v46
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v14
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v16
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v20
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v22
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v3
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v4
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v6
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v7
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v8
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v9
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v10
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v11
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v15
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v17
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v19
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v21
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v23
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v24
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v25
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v26
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v27
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v28
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v29
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v30
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v31
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v33
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v35
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v36
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v37
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v39
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v44
-; GCN-NEXT: .LBB46_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v3, v1
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v0, v50, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 4, v50
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v45, v1, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v44, v1, v0
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v47, v1, v0
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v46, v1, v0
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 20, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 28, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 36, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 40, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 44, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v63, v1, v0
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 48, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v60, v1, v0
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 52, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v30, v1, v0
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 56, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v61, v1, v0
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v7, v1, v0
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 64, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v62, v1, v0
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x44, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v5, v1, v0
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x48, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v52, v12, v0
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x4c, v50
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v54, v12, v0
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x50, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v32
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v1, v11, v0
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x54, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v20
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v11, v11, v0
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x58, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v26
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v15, v12, v0
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x5c, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v19, v6, v0
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x60, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v9, v6, v0
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 0x64, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v13, v2, v0
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x68, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v17, v2, v0
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x6c, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v21, v2, v0
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x70, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v25, v2, v0
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x74, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_or_b32_e32 v29, v2, v0
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x78, v50
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v50
-; GCN-NEXT: buffer_store_dword v45, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v56, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v48, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v8, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v14, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v18, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v22, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v60, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v62, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v23
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v27
-; VI-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v38
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v48
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v50
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB46_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: .LBB46_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB46_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v18, 0x300
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v0, v2, v0
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v2, v3, v2
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v5, v4
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v6, v5
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v6, v7, v6
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v35
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v36
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 3, v62
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v8, v9, v8
-; VI-NEXT: v_add_u16_e32 v9, 3, v63
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v59
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: v_add_u16_e32 v10, 3, v60
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v10, v11, v10
-; VI-NEXT: v_add_u16_e32 v11, 3, v58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v11, v12, v11
-; VI-NEXT: v_add_u16_e32 v12, 3, v47
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v46
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v12, v13, v12
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v13, v14, v13
-; VI-NEXT: v_add_u16_e32 v14, 3, v43
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v15, 3, v42
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: v_or_b32_e32 v14, v15, v14
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_e32 v16, v19, v16
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v40
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v55
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v21, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v48, 3, v48
-; VI-NEXT: v_or_b32_sdwa v48, v49, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v49, 3, v49
-; VI-NEXT: v_or_b32_sdwa v49, v50, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v20, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v50, 3, v50
-; VI-NEXT: v_or_b32_sdwa v50, v51, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v39
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v51, 3, v51
-; VI-NEXT: v_or_b32_sdwa v51, v52, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v52, 3, v52
-; VI-NEXT: v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v53, 3, v53
-; VI-NEXT: v_or_b32_sdwa v53, v54, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v54, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v19, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v39, v18
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v52
-; VI-NEXT: v_or_b32_e32 v19, v39, v19
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v50
-; VI-NEXT: v_or_b32_e32 v20, v39, v20
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v48
-; VI-NEXT: v_or_b32_e32 v21, v39, v21
-; VI-NEXT: v_or_b32_e32 v31, v31, v54
-; VI-NEXT: .LBB46_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v25
-; GFX9-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v39
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v49
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v52
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v3
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB46_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v17, v16, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v17, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v18, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v19, v20, v19, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v20, v21, v20, s6
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v28, v29, v28, s6
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: .LBB46_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB46_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_perm_b32 v0, v2, v0, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_perm_b32 v1, v1, v4, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v38, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_e32 v35, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v22
-; GFX9-NEXT: v_add_u16_e32 v36, 0x300, v36
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v33, 0x300, v21
-; GFX9-NEXT: v_add_u16_e32 v34, 0x300, v23
-; GFX9-NEXT: v_perm_b32 v29, v34, v29, s6
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v39
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v43
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v16
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v18
-; GFX9-NEXT: v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v20
-; GFX9-NEXT: v_perm_b32 v30, v33, v30, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v39, 0x300, v50
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v51
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v32, 0x300, v19
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v53
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v55
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v50, 0x300, v40
-; GFX9-NEXT: v_perm_b32 v21, v50, v21, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v41
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v51, 0x300, v42
-; GFX9-NEXT: v_perm_b32 v20, v51, v20, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v45
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v24
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v26
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v37
-; GFX9-NEXT: v_add_u16_e32 v37, 0x300, v38
-; GFX9-NEXT: v_add_u16_e32 v38, 0x300, v48
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v49
-; GFX9-NEXT: v_add_u16_e32 v48, 0x300, v52
-; GFX9-NEXT: v_add_u16_e32 v49, 0x300, v54
-; GFX9-NEXT: v_add_u16_e32 v52, 0x300, v44
-; GFX9-NEXT: v_add_u16_e32 v53, 0x300, v46
-; GFX9-NEXT: v_perm_b32 v18, v53, v18, s6
-; GFX9-NEXT: v_perm_b32 v19, v52, v19, s6
-; GFX9-NEXT: v_perm_b32 v22, v49, v22, s6
-; GFX9-NEXT: v_perm_b32 v23, v48, v23, s6
-; GFX9-NEXT: v_perm_b32 v24, v39, v24, s6
-; GFX9-NEXT: v_perm_b32 v25, v38, v25, s6
-; GFX9-NEXT: v_perm_b32 v26, v37, v26, s6
-; GFX9-NEXT: v_perm_b32 v27, v36, v27, s6
-; GFX9-NEXT: v_perm_b32 v28, v35, v28, s6
-; GFX9-NEXT: .LBB46_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:468
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:456
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v37, v30 :: v_dual_mov_b32 v54, v24
-; GFX11-NEXT: v_dual_mov_b32 v55, v28 :: v_dual_mov_b32 v52, v26
-; GFX11-NEXT: v_dual_mov_b32 v50, v22 :: v_dual_mov_b32 v53, v20
-; GFX11-NEXT: v_dual_mov_b32 v34, v18 :: v_dual_mov_b32 v51, v16
-; GFX11-NEXT: v_dual_mov_b32 v36, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v39, v4
-; GFX11-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v33, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v94, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v88, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v93, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v91, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v150, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v182, off, s32
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v45, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v46, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v47, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v61, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v62, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v63, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v72, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v73, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v79, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v89, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v90, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v95, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v104, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v105, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v132, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v161, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v160, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v176, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v167, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(62)
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v0
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v14
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v26
-; GFX11-NEXT: s_waitcnt vmcnt(61)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v88
-; GFX11-NEXT: s_waitcnt vmcnt(59)
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v93
-; GFX11-NEXT: s_waitcnt vmcnt(57)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v91
-; GFX11-NEXT: s_waitcnt vmcnt(55)
-; GFX11-NEXT: v_lshlrev_b16 v92, 8, v92
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v150
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v150, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v41, 8, v40
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v40, 8, v43
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v43, 8, v44
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v45
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v46
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v47
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v56
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v58
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v58, 8, v59
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v44, 8, v60
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v60, 8, v61
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v59, 8, v62
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v63
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v72
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v73
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v74
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v74, 8, v75
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v76
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v75, 8, v77
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v78
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v78, 8, v79
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v77, 8, v89
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v90
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v76, 8, v95
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v104
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v105
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v94
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v28
-; GFX11-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB46_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v48
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v70
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v71
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v84
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v82
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v51
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v34
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v66
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v128
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v113
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v132
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v100
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v161
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v160
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v176
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v167
-; GFX11-NEXT: v_perm_b32 v2, v2, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v5, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v7, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v9, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v11, v10, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v102
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v114
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v117
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v130
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v181
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v150
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v41
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v40
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v43
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v182
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v46
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v45
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v57
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v56
-; GFX11-NEXT: v_perm_b32 v7, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v12, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v14, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v16, v15, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v147
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v119
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v144
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v162
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v146
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v164
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v148
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v58
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v44
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v60
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v59
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v62
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v47
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v72
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v63
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v74
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v73
-; GFX11-NEXT: v_perm_b32 v12, v13, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v15, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v17, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v19, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v21, v20, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v166
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v145
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v163
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v179
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v165
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v183
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v180
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v65
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v75
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v61
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v78
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v77
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v79
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v76
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v90
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v89
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v92
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v91
-; GFX11-NEXT: v_perm_b32 v17, v18, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v20, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v22, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v24, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v26, v25, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v101
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v93
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v88
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v104
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v95
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v105
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v94
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v108
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v107
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v110
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v109
-; GFX11-NEXT: v_perm_b32 v22, v23, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v25, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v27, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v29, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v31, v30, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v103
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v99
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v129
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v131
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v116
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v111
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v106
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v122
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v121
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v123
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v120
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v125
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v124
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v126
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v127
-; GFX11-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v30, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v32, v31, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v34, v33, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v36, v35, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: .LBB46_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB46_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v118, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v131, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v116, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v129, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v35, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v126, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v127, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v125, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v124, v3
-; GFX11-NEXT: v_add_nc_u16 v33, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v116, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v98, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v1, v112, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_add_nc_u16 v3, v99, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v103, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v120, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v122, v1
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v121, v3
-; GFX11-NEXT: v_add_nc_u16 v99, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v111, v4
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v81, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v101, 3
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v86, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v97, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v83, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v106, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v110, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v109, v0
-; GFX11-NEXT: v_add_nc_u16 v83, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v108, v3
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v107, v4
-; GFX11-NEXT: v_add_nc_u16 v86, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v80, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v69, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v105, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v94, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v104, v2
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v95, v3
-; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v93, v4
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(26)
-; GFX11-NEXT: v_add_nc_u16 v0, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v65, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(24)
-; GFX11-NEXT: v_add_nc_u16 v3, v183, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v180, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v88, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v92, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v91, v1
-; GFX11-NEXT: v_add_nc_u16 v65, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v90, v3
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v89, v4
-; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_add_nc_u16 v1, v179, 3
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v165, 3
-; GFX11-NEXT: v_add_nc_u16 v80, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_add_nc_u16 v0, v177, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v163, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_add_nc_u16 v4, v166, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v79, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v76, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v78, v0
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v77, v3
-; GFX11-NEXT: v_add_nc_u16 v85, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v75, v4
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v145, 3
-; GFX11-NEXT: v_add_nc_u16 v97, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_add_nc_u16 v1, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v148, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_add_nc_u16 v3, v178, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v164, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v61, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v74, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v73, v2
-; GFX11-NEXT: v_add_nc_u16 v101, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v72, v3
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v63, v4
-; GFX11-NEXT: v_add_nc_u16 v103, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_add_nc_u16 v2, v162, 3
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v146, 3
-; GFX11-NEXT: v_add_nc_u16 v112, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(10)
-; GFX11-NEXT: v_add_nc_u16 v1, v149, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v144, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_add_nc_u16 v4, v147, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v62, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v47, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v60, v1
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v59, v3
-; GFX11-NEXT: v_add_nc_u16 v118, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v58, v4
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v119, 3
-; GFX11-NEXT: v_add_nc_u16 v119, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_add_nc_u16 v2, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v130, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_add_nc_u16 v3, v133, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v117, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v44, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v57, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v56, v0
-; GFX11-NEXT: v_add_nc_u16 v117, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v46, v3
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v45, v4
-; GFX11-NEXT: v_add_nc_u16 v129, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_add_nc_u16 v0, v114, 3
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v96, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v2, v102, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v87, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v55, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v43, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v182, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v41, v2
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v40, v3
-; GFX11-NEXT: v_add_nc_u16 v55, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v181, v4
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v37, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v52, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v53, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v50, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v150, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v176, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v167, v1
-; GFX11-NEXT: v_add_nc_u16 v50, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v161, v3
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v160, v4
-; GFX11-NEXT: v_add_nc_u16 v52, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v49, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v36, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v132, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v128, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v115, v0
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v100, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v113, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v39, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v32, 3
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_or_b32_e32 v35, v71, v35
-; GFX11-NEXT: v_or_b32_e32 v33, v82, v33
-; GFX11-NEXT: v_or_b32_e32 v0, v84, v0
-; GFX11-NEXT: v_or_b32_e32 v36, v70, v36
-; GFX11-NEXT: v_or_b32_e32 v32, v66, v32
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v38, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v36
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v39, 0x300, v1
-; GFX11-NEXT: v_perm_b32 v1, v33, v38, 0x5040100
-; GFX11-NEXT: v_perm_b32 v0, v35, v0, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v32, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v36, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v39, v34, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v52, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v50, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v96, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v129, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v12, v117, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v119, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v118, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v112, v15, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v103, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v17, v101, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v97, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v85, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v80, v20, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v69, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v22, v65, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v64, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v68, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v67, v25, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v86, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v27, v83, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v81, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v99, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v98, v30, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v116, v31, 0x5040100
-; GFX11-NEXT: .LBB46_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:516
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:580
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v11
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v16
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v20
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v24
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v23
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v25
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v28
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v27
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v29
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v44
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v37
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v36
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v62
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v53
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v61
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v58
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v59
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v56
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v57
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v43
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v42
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v41
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v40
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v55
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v54
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v52
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v51
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v50
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v38
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v35
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v8
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v45
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v46
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v47
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v14
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v15
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v16
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v31
-; GCN-NEXT: v_or_b32_e32 v61, v32, v14
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v35
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v15, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v55, v5, v6
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v59
-; GCN-NEXT: v_or_b32_e32 v43, v7, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v63
-; GCN-NEXT: v_or_b32_e32 v41, v9, v5
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v23
-; GCN-NEXT: v_or_b32_e32 v47, v10, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v2, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v32
-; GCN-NEXT: v_or_b32_e32 v58, v11, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v3
-; GCN-NEXT: v_or_b32_e32 v54, v4, v1
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_or_b32_e32 v60, v12, v1
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v24
-; GCN-NEXT: v_or_b32_e32 v44, v13, v2
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v30
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v5, v1
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v34
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v45, v6, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v36
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v7, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v37
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v9, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v10, v1
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v48
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v56, v11, v5
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v49
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v12, v3
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v50
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v57, v13, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v52
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v28, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v33, v7
-; GCN-NEXT: v_bfe_u32 v7, v35, 8, 8
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v7, v4
-; GCN-NEXT: v_bfe_u32 v4, v59, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v59, v4, v9
-; GCN-NEXT: v_bfe_u32 v4, v23, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v4, v1
-; GCN-NEXT: v_bfe_u32 v1, v32, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v1, v10
-; GCN-NEXT: v_bfe_u32 v1, v31, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v1, v5
-; GCN-NEXT: v_bfe_u32 v1, v22, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v1, v11
-; GCN-NEXT: v_bfe_u32 v1, v16, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v1, v3
-; GCN-NEXT: v_bfe_u32 v1, v15, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v15, v8, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v21, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v1, v6
-; GCN-NEXT: v_bfe_u32 v1, v20, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v16, v38, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v18, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v29, v2
-; GCN-NEXT: v_bfe_u32 v2, v14, 8, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v2, v24, 8, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v2, v30, 8, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v2, v19, 8, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v62, v61, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v62, v61, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v62, v61, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v43, v55, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v43, v55, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v43, v55, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v47, v41, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v47, v41, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v47, v41, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v58, v40, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v58, v40, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v58, v40, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v60, v54, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v60, v54, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v60, v54, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v51, v44, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v51, v44, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v51, v44, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v25, v45, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v25, v45, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v25, v45, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v26, v46, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v26, v46, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v26, v46, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v27, v56, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v27, v56, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v27, v56, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v28, v57, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v28, v57, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v28, v57, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v35, v42, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v35, v42, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v35, v42, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v23, v59, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v23, v59, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v23, v59, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v9, v7, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v9, v7, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v2, v9, v7, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v3, v15, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v3, v15, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v3, v15, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v1, v16, 24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v1, v16, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v2, v1, v16, 8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v62
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v43
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v47
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v58
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v60
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v51
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v27
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v23
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v9
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v5
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v53, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 8, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v14, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v17, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; kill: killed $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; kill: killed $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: .LBB47_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v38
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v29
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v15
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v15, v16
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v8, v15
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v19
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v8
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v14
-; GCN-NEXT: v_or_b32_e32 v8, v8, v17
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v49
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v16
-; GCN-NEXT: v_or_b32_e32 v8, v8, v19
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v48
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v19, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v18
-; GCN-NEXT: v_or_b32_e32 v8, v8, v21
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v39
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v14
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v54, 0x38000000, v54
-; GCN-NEXT: v_add_f32_e32 v57, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v55, 0x38000000, v55
-; GCN-NEXT: v_add_f32_e32 v40, 0x38000000, v40
-; GCN-NEXT: v_add_f32_e32 v41, 0x38000000, v41
-; GCN-NEXT: v_add_f32_e32 v58, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v59, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v42, 0x38000000, v42
-; GCN-NEXT: v_add_f32_e32 v43, 0x38000000, v43
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v44, 0x38000000, v44
-; GCN-NEXT: v_add_f32_e32 v60, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v45, 0x38000000, v45
-; GCN-NEXT: v_add_f32_e32 v61, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v46, 0x38000000, v46
-; GCN-NEXT: v_add_f32_e32 v62, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v47, 0x38000000, v47
-; GCN-NEXT: v_add_f32_e32 v63, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v56, 0x38000000, v56
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v54
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v57
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v55
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v40
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v58
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v59
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v43
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v60
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v45
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v62
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v63
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v53
-; GCN-NEXT: v_mov_b32_e32 v53, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v56
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v33
-; GCN-NEXT: v_mov_b32_e32 v33, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v17
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v7
-; GCN-NEXT: v_or_b32_e32 v59, v23, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v8
-; GCN-NEXT: v_or_b32_e32 v23, v19, v5
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_or_b32_e32 v42, v9, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v38
-; GCN-NEXT: v_or_b32_e32 v35, v35, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: v_or_b32_e32 v57, v28, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v51
-; GCN-NEXT: v_or_b32_e32 v28, v27, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v36
-; GCN-NEXT: v_or_b32_e32 v56, v13, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v29
-; GCN-NEXT: v_or_b32_e32 v27, v22, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v30
-; GCN-NEXT: v_or_b32_e32 v46, v24, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v32
-; GCN-NEXT: v_or_b32_e32 v26, v26, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v60
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v45, v34, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v37
-; GCN-NEXT: v_or_b32_e32 v25, v25, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v62
-; GCN-NEXT: v_or_b32_e32 v44, v12, v14
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_bfe_u32 v5, v62, 8, 8
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v51, v40, v38
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v5, v18, 8, 8
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v54, v54, v13
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v5, v60, 8, 8
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_or_b32_e32 v60, v52, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v1, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v40, v50, v22
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v58, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_or_b32_e32 v58, v43, v30
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v41, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_or_b32_e32 v41, v61, v24
-; GCN-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v55, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v47, v47, v32
-; GCN-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v49, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v55, v4, v31
-; GCN-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v48, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v43, v3, v36
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v11, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v61, v2, v34
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v21, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v62, v15, v37
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v20, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v53, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_bfe_u32 v1, v33, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14) expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v19, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v62, v61, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v62, v61, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v62, v61, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v55, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v55, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v33, v43, v55, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v41, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v41, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v41, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v58, v40, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v58, v40, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v58, v40, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v60, v54, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v60, v54, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v60, v54, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v44, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v44, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v44, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v25, v45, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v25, v45, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v25, v45, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v26, v46, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v26, v46, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v26, v46, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v27, v56, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v27, v56, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v27, v56, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v28, v57, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v28, v57, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v28, v57, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v35, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v23, v59, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v23, v59, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v23, v59, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 8
-; GCN-NEXT: v_mov_b32_e32 v5, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v3, 8
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v3, v15, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v3, v15, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v3, v15, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v16, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v16, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v16, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v58
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v53, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v14, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v17, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB47_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v61
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v62
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v2, v4
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v33
-; GCN-NEXT: v_or_b32_e32 v29, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v43
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v30, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v41
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v2, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v47
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v61, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v40
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v62, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v58
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v44
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v51
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v45
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v25
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v46
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v56
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v27
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v22, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v57
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v28
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v25, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v42
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v26, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v27, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v28, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_mov_b32_e32 v5, v19
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v17
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v15
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v53
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v18, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v16
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v31, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v49, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v4, v1, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v5, v1, v33
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v61
-; GCN-NEXT: v_or_b32_e32 v61, v3, v34
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v62
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v36
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v29, v37
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v30, v38
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v48
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v9, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v11, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v12, v12, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v13, v13, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v22, v22, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v25, v25, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v42
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x50, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v27, v27, v43
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v23, v23, v44
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v19, v19, v45
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v21, v21, v46
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v17, v17, v47
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v20, v20, v56
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v15, v15, v57
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v58
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v59
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v16, v60
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v13
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v9
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v5
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v3
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v1
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v30
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v29
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v27
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v25
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v62, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v28
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v26
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v57, 16, v24
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_lshrrev_b32_e32 v59, 16, v22
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v21
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v20
-; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v17
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v61
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v60
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB47_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v16
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v14
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v13
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v12
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v11
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v10
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v9
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v8
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v7
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v4
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v4
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v3
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v2
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v2
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v1
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v61
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v61
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v60
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v30
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v30
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v29
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v28
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v27
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v26
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v26
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v25
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v24
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v24
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v23
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v22
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v22
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v21
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v20
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v20
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v19
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v18
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v17
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v31, v33
-; VI-NEXT: v_mov_b32_e32 v33, v43
-; VI-NEXT: v_lshrrev_b64 v[42:43], 24, v[5:6]
-; VI-NEXT: v_mov_b32_e32 v43, v33
-; VI-NEXT: v_mov_b32_e32 v33, v46
-; VI-NEXT: v_lshrrev_b64 v[45:46], 24, v[3:4]
-; VI-NEXT: v_mov_b32_e32 v46, v33
-; VI-NEXT: v_mov_b32_e32 v33, v53
-; VI-NEXT: v_lshrrev_b64 v[52:53], 24, v[1:2]
-; VI-NEXT: v_mov_b32_e32 v53, v33
-; VI-NEXT: v_lshrrev_b64 v[32:33], 24, v[60:61]
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[32:33], 24, v[29:30]
-; VI-NEXT: v_lshrrev_b64 v[34:35], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v33, v36
-; VI-NEXT: v_lshrrev_b64 v[35:36], 24, v[25:26]
-; VI-NEXT: v_lshrrev_b64 v[49:50], 24, v[23:24]
-; VI-NEXT: v_mov_b32_e32 v36, v33
-; VI-NEXT: v_mov_b32_e32 v33, v41
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[21:22]
-; VI-NEXT: v_mov_b32_e32 v34, v51
-; VI-NEXT: v_lshrrev_b64 v[50:51], 24, v[17:18]
-; VI-NEXT: v_mov_b32_e32 v41, v33
-; VI-NEXT: v_mov_b32_e32 v33, v31
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[19:20]
-; VI-NEXT: v_mov_b32_e32 v51, v34
-; VI-NEXT: .LBB47_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB47_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v63, 0x200
-; VI-NEXT: v_add_f16_sdwa v31, v18, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_or_b32_e32 v32, v18, v31
-; VI-NEXT: v_add_f16_sdwa v31, v17, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_or_b32_e32 v31, v17, v31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_sdwa v31, v20, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
-; VI-NEXT: v_or_b32_e32 v32, v20, v31
-; VI-NEXT: v_add_f16_sdwa v31, v19, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_or_b32_e32 v31, v19, v31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_sdwa v34, v22, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v34
-; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
-; VI-NEXT: v_or_b32_e32 v32, v22, v31
-; VI-NEXT: v_add_f16_sdwa v31, v21, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
-; VI-NEXT: v_or_b32_e32 v31, v21, v31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_sdwa v31, v24, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
-; VI-NEXT: v_or_b32_e32 v32, v24, v31
-; VI-NEXT: v_add_f16_sdwa v31, v23, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
-; VI-NEXT: v_or_b32_e32 v31, v23, v31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_sdwa v31, v26, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
-; VI-NEXT: v_or_b32_e32 v36, v26, v31
-; VI-NEXT: v_add_f16_sdwa v31, v25, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
-; VI-NEXT: v_or_b32_e32 v35, v25, v31
-; VI-NEXT: v_add_f16_sdwa v31, v28, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
-; VI-NEXT: v_or_b32_e32 v38, v28, v31
-; VI-NEXT: v_add_f16_sdwa v31, v27, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
-; VI-NEXT: v_or_b32_e32 v37, v27, v31
-; VI-NEXT: v_add_f16_sdwa v31, v30, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
-; VI-NEXT: v_add_f16_sdwa v32, v29, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
-; VI-NEXT: v_or_b32_e32 v49, v30, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_or_b32_e32 v48, v29, v31
-; VI-NEXT: v_add_f16_sdwa v31, v61, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_e32 v61, 0x200, v61
-; VI-NEXT: v_add_f16_sdwa v32, v60, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v60, 0x200, v60
-; VI-NEXT: v_or_b32_e32 v51, v61, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_or_b32_e32 v50, v60, v31
-; VI-NEXT: v_add_f16_sdwa v31, v2, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v32, v1, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v53, v2, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_or_b32_e32 v52, v1, v31
-; VI-NEXT: v_add_f16_sdwa v31, v4, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v32, v3, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v46, v4, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_or_b32_e32 v45, v3, v31
-; VI-NEXT: v_add_f16_sdwa v31, v6, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v32, v5, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v43, v6, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_add_f16_sdwa v44, v8, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v42, v5, v31
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v32, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v44
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v41, v8, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_or_b32_e32 v40, v7, v31
-; VI-NEXT: v_add_f16_sdwa v31, v10, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_add_f16_sdwa v32, v9, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v55, v10, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v32
-; VI-NEXT: v_add_f16_sdwa v39, v12, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v54, v9, v31
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_add_f16_sdwa v33, v11, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v39
-; VI-NEXT: v_add_f16_sdwa v47, v14, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v32, v12, v31
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v33
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_add_f16_sdwa v33, v13, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v56, 16, v47
-; VI-NEXT: v_or_b32_e32 v57, v14, v56
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v56, 16, v33
-; VI-NEXT: v_add_f16_sdwa v33, v16, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_add_f16_sdwa v63, v15, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v58, 16, v33
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v59, v16, v58
-; VI-NEXT: v_lshlrev_b32_e32 v58, 16, v63
-; VI-NEXT: v_or_b32_e32 v58, v15, v58
-; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v59
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v58
-; VI-NEXT: v_lshrrev_b64 v[58:59], 24, v[58:59]
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v56, v13, v56
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v57
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v56
-; VI-NEXT: v_lshrrev_b64 v[56:57], 24, v[56:57]
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v31, v11, v31
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v32
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v31
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[31:32]
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v55
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v54
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[54:55]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v41
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v40
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[40:41]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v43
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v42
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v46
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v45
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v53
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v52
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v51
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v50
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[50:51]
-; VI-NEXT: v_lshrrev_b64 v[42:43], 24, v[42:43]
-; VI-NEXT: v_lshrrev_b64 v[45:46], 24, v[45:46]
-; VI-NEXT: v_lshrrev_b64 v[52:53], 24, v[52:53]
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v49
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v48
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[48:49]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v38
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v37
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[37:38]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v36
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v35
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[35:36], 24, v[35:36]
-; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v54, v39
-; VI-NEXT: v_mov_b32_e32 v37, v44
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_mov_b32_e32 v56, v58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v49
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v48
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[49:50], 24, v[48:49]
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v48, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v51
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v50
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[50:51]
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v51
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v50
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[50:51]
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshrrev_b32_e32 v32, 8, v51
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 8, v50
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v33, 8, 8
-; VI-NEXT: v_mov_b32_e32 v33, v47
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v33, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v39, 8, 8
-; VI-NEXT: v_mov_b32_e32 v39, v63
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[50:51], 24, v[50:51]
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_bfe_u32 v32, v63, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v44, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_bfe_u32 v32, v47, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v44, v32
-; VI-NEXT: v_bfe_u32 v32, v32, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v55, v32
-; VI-NEXT: v_bfe_u32 v32, v32, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v36, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_bfe_u32 v32, v32, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v58, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v57, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v59, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v34, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v53, 8, 8
-; VI-NEXT: v_mov_b32_e32 v58, v57
-; VI-NEXT: v_mov_b32_e32 v57, v59
-; VI-NEXT: v_mov_b32_e32 v59, v34
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v32, v41, 8, 8
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: .LBB47_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v1, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v32
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v2, v2, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v34, v32, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v7, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v37, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v63, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v62, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v11, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v54, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v13, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v15, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_or_b32_sdwa v1, v43, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v40
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v59, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v49
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v35
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v56, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB47_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(45)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(46)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[21:22]
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[19:20]
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 24, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v17
-; GFX9-NEXT: v_lshrrev_b64 v[42:43], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; GFX9-NEXT: .LBB47_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB47_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_pk_add_f16 v32, v32, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: v_lshrrev_b64 v[42:43], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 24, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v17
-; GFX9-NEXT: .LBB47_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v63
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v61
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v49
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v42
-; GFX9-NEXT: v_or_b32_sdwa v34, v58, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v57, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v50
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v53
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v56
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB47_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-NEXT: .LBB47_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB47_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-NEXT: .LBB47_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v74
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v60
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v47
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v63
-; GFX11-NEXT: v_or_b32_e32 v54, v67, v54
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v62
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v64
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v45
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v43
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v51
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v67
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v181
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v55
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v167
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v55, v64
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v164
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v54
-; GFX11-NEXT: v_or_b32_e32 v52, v55, v52
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v150
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v53
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v52, v54, v55
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v145
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v144
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v51, v54, v51
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v52
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v51, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v131
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v119
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v117
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v116
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v115
-; GFX11-NEXT: v_or_b32_e32 v51, v51, v52
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v53
-; GFX11-NEXT: v_or_b32_e32 v50, v54, v50
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v51
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v50
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v52
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v102
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v86
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v68
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v73
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v66
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v48
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v61
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v56
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v65
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v44
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v40
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v182
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v176
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v160
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v146
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v135
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v132
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v103
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v101
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v99
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v81
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v69
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v128i8_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v49, v7
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v55, v1
-; GCN-NEXT: v_mov_b32_e32 v60, v0
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:56
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:48
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v14
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 24, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v22
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 24, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v30
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 24, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v10
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v18
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:392
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:112
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v23
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v21
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v19
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v29
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v9
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v13
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:156
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:152
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:148
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:144
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:188
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:184
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:180
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:176
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:252
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:248
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:244
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:240
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:284
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:280
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:276
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:272
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:316
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:312
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:308
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:304
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:348
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:344
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:340
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:336
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:388
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:380
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:376
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:372
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:368
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 8, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v5
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:300
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:328
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:352
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:364
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:360
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:384
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v4
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; kill: killed $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; kill: killed $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; kill: killed $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; kill: killed $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; kill: killed $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; kill: killed $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; kill: killed $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; kill: killed $vgpr25
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v4, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v6, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v7, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v8, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v32, v1, v24
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v28, v1, v28
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v30
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v1
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v17
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v21, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v25, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v25, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v25, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v25, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v25, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v25, v1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v25, v45
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v13, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v13, v51
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v12, v38
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v12, v39
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v56, v48, v12
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v50, v12
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v2, v36, v13
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v53, v13
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v3, v40, v22
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v47, v22
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v1, v26
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v1, v63, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v1, v33
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v1, v38
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v1, v48
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v59, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v58, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v45, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v48, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v43, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v41, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v47, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v1, v53
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v54, v1, v54
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v30
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v1
-; GCN-NEXT: v_mov_b32_e32 v1, v56
-; GCN-NEXT: v_or_b32_e32 v8, v49, v1
-; GCN-NEXT: v_or_b32_e32 v31, v51, v2
-; GCN-NEXT: v_or_b32_e32 v56, v57, v3
-; GCN-NEXT: v_or_b32_e32 v4, v55, v0
-; GCN-NEXT: v_or_b32_e32 v5, v40, v62
-; GCN-NEXT: v_or_b32_e32 v6, v37, v17
-; GCN-NEXT: v_or_b32_e32 v7, v25, v61
-; GCN-NEXT: v_or_b32_e32 v37, v34, v59
-; GCN-NEXT: v_or_b32_e32 v25, v9, v63
-; GCN-NEXT: v_or_b32_e32 v38, v10, v44
-; GCN-NEXT: v_or_b32_e32 v51, v11, v52
-; GCN-NEXT: v_or_b32_e32 v55, v12, v50
-; GCN-NEXT: v_or_b32_e32 v49, v13, v48
-; GCN-NEXT: v_or_b32_e32 v40, v14, v41
-; GCN-NEXT: v_or_b32_e32 v11, v15, v39
-; GCN-NEXT: v_or_b32_e32 v57, v16, v53
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v36, v12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v18, v14
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v19, v13
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v20, v34
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v21, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v22, v21
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v23, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v24, v23
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v33, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v26, v58
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v27, v46
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v28, v45
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v29, v43
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v30, v42
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v35, v47
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v54
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v1
-; GCN-NEXT: v_mov_b32_e32 v1, v8
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v2
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v0
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v62
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v17
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v61
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v59
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v63
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v44
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v52
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v50
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v48
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v41
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v39
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v61, v1, v12, 16
-; GCN-NEXT: v_mov_b32_e32 v12, v33
-; GCN-NEXT: v_alignbit_b32 v33, v31, v14, 16
-; GCN-NEXT: v_mov_b32_e32 v27, v56
-; GCN-NEXT: v_alignbit_b32 v59, v56, v13, 16
-; GCN-NEXT: v_mov_b32_e32 v13, v19
-; GCN-NEXT: v_alignbit_b32 v29, v4, v34, 16
-; GCN-NEXT: v_alignbit_b32 v0, v5, v20, 16
-; GCN-NEXT: v_alignbit_b32 v14, v6, v21, 16
-; GCN-NEXT: v_mov_b32_e32 v21, v18
-; GCN-NEXT: v_mov_b32_e32 v18, v10
-; GCN-NEXT: v_mov_b32_e32 v56, v16
-; GCN-NEXT: v_mov_b32_e32 v16, v36
-; GCN-NEXT: v_mov_b32_e32 v10, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v7, v22, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v62, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v37, v23, 16
-; GCN-NEXT: v_mov_b32_e32 v23, v9
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v25, v24, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v10, v58, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v34, v51
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v51, v46, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v58, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v55, v45, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v52, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v49, v43, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v46, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v15, v40, v42, 16
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v41, v11
-; GCN-NEXT: v_alignbit_b32 v11, v11, v47, 16
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v3, v57, v54, 16
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v53
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; kill: killed $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; kill: killed $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; kill: killed $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; kill: killed $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; kill: killed $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: .LBB48_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v25, v27
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v26
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v30, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v38, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_or_b32_e32 v1, v20, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v1, v51, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v1, v28, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v6, v22, v7
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v1, v8
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v45, v9
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v9, v24, v10
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v1, v13
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v1, v14
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v1, v18
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v1, v19
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v1, v20
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v1, v21
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v1, v22
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_mov_b32_e32 v2, v35
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v51, v1, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v17
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_mov_b32_e32 v4, v49
-; GCN-NEXT: v_mov_b32_e32 v43, v42
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_mov_b32_e32 v44, v40
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v57, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v1, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v59, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v61, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v1, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v23
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v24, v23
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v24, v23
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v24, v23
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v24, v23
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_mov_b32_e32 v52, v36
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v24, v23
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_mov_b32_e32 v45, v53
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v53, v24, v23
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v24, v23
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v27, v25
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v31, v29
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v17
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v17
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v36, v34
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 3, v11
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v2, v37
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v49, v2, v49
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v63, v54
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v43
-; GCN-NEXT: v_and_b32_e32 v40, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v2, v40
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_or_b32_e32 v41, v47, v41
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_or_b32_e32 v43, v44, v43
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: v_or_b32_e32 v44, v45, v44
-; GCN-NEXT: v_add_i32_e32 v45, vcc, 3, v15
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_or_b32_e32 v45, v52, v45
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_or_b32_e32 v47, v50, v47
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 3, v4
-; GCN-NEXT: v_and_b32_e32 v56, 0xff, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v56
-; GCN-NEXT: v_or_b32_e32 v56, v48, v56
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x300, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v61, v61, v2
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v1, v4
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s7, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v20
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v21
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s7, v26
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s7, v30
-; GCN-NEXT: v_add_i32_e32 v30, vcc, s7, v35
-; GCN-NEXT: v_add_i32_e32 v35, vcc, s7, v51
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s7, v42
-; GCN-NEXT: v_add_i32_e32 v42, vcc, s7, v46
-; GCN-NEXT: v_add_i32_e32 v46, vcc, s7, v57
-; GCN-NEXT: v_add_i32_e32 v57, vcc, s7, v38
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s7, v59
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s7, v62
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v51
-; GCN-NEXT: v_and_b32_e32 v42, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v57
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v59
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v38, v1
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v38, v6
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v38, v7
-; GCN-NEXT: v_or_b32_e32 v8, v39, v8
-; GCN-NEXT: v_or_b32_e32 v9, v53, v9
-; GCN-NEXT: v_or_b32_e32 v10, v55, v10
-; GCN-NEXT: v_or_b32_e32 v11, v23, v11
-; GCN-NEXT: v_or_b32_e32 v12, v24, v12
-; GCN-NEXT: v_or_b32_e32 v13, v25, v13
-; GCN-NEXT: v_or_b32_e32 v14, v27, v14
-; GCN-NEXT: v_or_b32_e32 v15, v28, v15
-; GCN-NEXT: v_or_b32_e32 v16, v29, v16
-; GCN-NEXT: v_or_b32_e32 v17, v31, v17
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_or_b32_e32 v19, v33, v19
-; GCN-NEXT: v_or_b32_e32 v20, v34, v20
-; GCN-NEXT: v_or_b32_e32 v21, v36, v21
-; GCN-NEXT: v_or_b32_e32 v22, v37, v22
-; GCN-NEXT: v_or_b32_e32 v24, v49, v26
-; GCN-NEXT: v_or_b32_e32 v25, v54, v30
-; GCN-NEXT: v_or_b32_e32 v26, v40, v35
-; GCN-NEXT: v_or_b32_e32 v28, v41, v51
-; GCN-NEXT: v_or_b32_e32 v30, v43, v42
-; GCN-NEXT: v_or_b32_e32 v33, v44, v46
-; GCN-NEXT: v_or_b32_e32 v34, v45, v57
-; GCN-NEXT: v_or_b32_e32 v38, v47, v58
-; GCN-NEXT: v_or_b32_e32 v39, v56, v59
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v61
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v49, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v35, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v53, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v30
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s6, v33
-; GCN-NEXT: v_add_i32_e32 v31, vcc, s6, v34
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v38
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v39
-; GCN-NEXT: v_alignbit_b32 v61, v1, v16, 16
-; GCN-NEXT: v_alignbit_b32 v33, v31, v56, 16
-; GCN-NEXT: v_alignbit_b32 v59, v25, v18, 16
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v29, v4, v8, 16
-; GCN-NEXT: v_alignbit_b32 v0, v5, v21, 16
-; GCN-NEXT: v_alignbit_b32 v2, v6, v13, 16
-; GCN-NEXT: v_alignbit_b32 v19, v7, v23, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v17, v14, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v14, v2
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v19, v15, v12, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v10, v53, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v11, v52, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v48, v51, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v36, v50, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v35, v49, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v3, v32, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v19, v27, v37, 16
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v1
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v31
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v4
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v5
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v6
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v7
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v62, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v37, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v10
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v34, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v58, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v48
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v52, v36
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v46, v35
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v41, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v27
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB48_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v61
-; GCN-NEXT: v_or_b32_e32 v8, v8, v19
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v19
-; GCN-NEXT: buffer_store_dword v8, v60, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 4, v60
-; GCN-NEXT: buffer_store_dword v1, v8, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v33
-; GCN-NEXT: v_or_b32_e32 v56, v1, v8
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: v_add_i32_e32 v45, vcc, 8, v60
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 12, v60
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 16, v60
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 20, v60
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 24, v60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v0
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v5
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 32, v60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v14
-; GCN-NEXT: v_or_b32_e32 v61, v2, v3
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 36, v60
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v47, v2, v3
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 40, v60
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 44, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v7
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v23, v3, v5
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 48, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v15, v3, v5
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v62
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v57, v3, v5
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 56, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v21, v3, v5
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v37
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v3, v5
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 64, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v25, v3, v5
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x44, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v10
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v27, v3, v5
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x48, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v30, v3, v5
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x4c, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v32, v3, v5
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x50, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v34, v3, v5
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x54, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v58
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v29, v3, v5
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x58, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v37, v3, v5
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x5c, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v52
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v3, v5
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v48, v3, v7
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v46
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v50, v3, v7
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v55, v3, v7
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x6c, v60
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v41
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v54, v3, v7
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x70, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v41, v3, v7
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x74, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v52, v3, v7
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x78, v60
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v60
-; GCN-NEXT: buffer_store_dword v56, v45, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v59, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v8, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v9, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v55, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v23
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v27
-; VI-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v38
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v48
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v50
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; VI-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; kill: killed $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; kill: killed $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: .LBB48_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v18, 0x300
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v29, 0x300, v29
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v0, v2, v0
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v2, v3, v2
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v5, v4
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v6, v5
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v28, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v27, 0x300, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v26, 0x300, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v34
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v6, v7, v6
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v35
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v36
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 3, v62
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v8, v9, v8
-; VI-NEXT: v_add_u16_e32 v9, 3, v63
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 3, v59
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: v_add_u16_e32 v10, 3, v60
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 3, v57
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v10, v11, v10
-; VI-NEXT: v_add_u16_e32 v11, 3, v58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 3, v56
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v11, v12, v11
-; VI-NEXT: v_add_u16_e32 v12, 3, v47
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 3, v46
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v12, v13, v12
-; VI-NEXT: v_add_u16_e32 v13, 3, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 3, v44
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v13, v14, v13
-; VI-NEXT: v_add_u16_e32 v14, 3, v43
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v15, 3, v42
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: v_or_b32_e32 v14, v15, v14
-; VI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v17, v17, v21
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_e32 v16, v19, v16
-; VI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v29, v40
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v30, v55
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v22, 0x300, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v21, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v48, 3, v48
-; VI-NEXT: v_or_b32_sdwa v48, v49, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v49, 3, v49
-; VI-NEXT: v_or_b32_sdwa v49, v50, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v20, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v50, 3, v50
-; VI-NEXT: v_or_b32_sdwa v50, v51, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v39, 3, v39
-; VI-NEXT: v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v39
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v51, 3, v51
-; VI-NEXT: v_or_b32_sdwa v51, v52, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v52, 3, v52
-; VI-NEXT: v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v53, 3, v53
-; VI-NEXT: v_or_b32_sdwa v53, v54, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v54, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v19, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v39, v18
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v52
-; VI-NEXT: v_or_b32_e32 v19, v39, v19
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v50
-; VI-NEXT: v_or_b32_e32 v20, v39, v20
-; VI-NEXT: v_add_u16_e32 v39, 0x300, v48
-; VI-NEXT: v_or_b32_e32 v21, v39, v21
-; VI-NEXT: v_or_b32_e32 v31, v31, v54
-; VI-NEXT: .LBB48_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:176
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT: s_waitcnt vmcnt(27)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v4
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v6
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v8
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v10
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v12
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v16
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v20
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(26)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v25
-; GFX9-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v37
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v39
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v49
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v52
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:156
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:172
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:180
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v53
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v42
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v43
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:192
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:188
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:196
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:204
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:212
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:240
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:220
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:228
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:236
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:244
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:280
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:252
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:260
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:268
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:276
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:304
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:284
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:292
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:300
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:308
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:336
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:344
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:316
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:324
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:332
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:340
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:352
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:360
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:368
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:376
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:348
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v3
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:356
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:364
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:372
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:384
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:380
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v17, v16, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v17, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v18, v19, v18, s6
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v19, v20, v19, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v20, v21, v20, s6
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v28, v29, v28, s6
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; kill: killed $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; kill: killed $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: .LBB48_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT: v_add_u16_e32 v24, 3, v24
-; GFX9-NEXT: v_perm_b32 v0, v2, v0, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_perm_b32 v1, v1, v4, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v25, 3, v25
-; GFX9-NEXT: v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v38, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_e32 v35, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v22
-; GFX9-NEXT: v_add_u16_e32 v36, 0x300, v36
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v23, 3, v23
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v33, 0x300, v21
-; GFX9-NEXT: v_add_u16_e32 v34, 0x300, v23
-; GFX9-NEXT: v_perm_b32 v29, v34, v29, s6
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v39
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v63
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v59
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v57
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v56
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v46
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v43
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v16
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v16
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v17, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v17
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v16
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v18
-; GFX9-NEXT: v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v19
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v20
-; GFX9-NEXT: v_perm_b32 v30, v33, v30, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v39, 0x300, v50
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v18
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v51
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v32, 0x300, v19
-; GFX9-NEXT: v_perm_b32 v31, v32, v31, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v53
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v55
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v50, 0x300, v40
-; GFX9-NEXT: v_perm_b32 v21, v50, v21, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v41
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v51, 0x300, v42
-; GFX9-NEXT: v_perm_b32 v20, v51, v20, s6
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v43
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v45
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v24
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v26
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v37
-; GFX9-NEXT: v_add_u16_e32 v37, 0x300, v38
-; GFX9-NEXT: v_add_u16_e32 v38, 0x300, v48
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v49
-; GFX9-NEXT: v_add_u16_e32 v48, 0x300, v52
-; GFX9-NEXT: v_add_u16_e32 v49, 0x300, v54
-; GFX9-NEXT: v_add_u16_e32 v52, 0x300, v44
-; GFX9-NEXT: v_add_u16_e32 v53, 0x300, v46
-; GFX9-NEXT: v_perm_b32 v18, v53, v18, s6
-; GFX9-NEXT: v_perm_b32 v19, v52, v19, s6
-; GFX9-NEXT: v_perm_b32 v22, v49, v22, s6
-; GFX9-NEXT: v_perm_b32 v23, v48, v23, s6
-; GFX9-NEXT: v_perm_b32 v24, v39, v24, s6
-; GFX9-NEXT: v_perm_b32 v25, v38, v25, s6
-; GFX9-NEXT: v_perm_b32 v26, v37, v26, s6
-; GFX9-NEXT: v_perm_b32 v27, v36, v27, s6
-; GFX9-NEXT: v_perm_b32 v28, v35, v28, s6
-; GFX9-NEXT: .LBB48_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v128i8_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:580
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:576
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:572
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:568
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:564
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:560
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:556
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:552
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:548
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:544
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:540
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:536
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:532
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:528
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:524
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:520
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:516
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:512
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:508
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:504
-; GFX11-NEXT: scratch_store_b32 off, v76, s32 offset:500
-; GFX11-NEXT: scratch_store_b32 off, v77, s32 offset:496
-; GFX11-NEXT: scratch_store_b32 off, v78, s32 offset:492
-; GFX11-NEXT: scratch_store_b32 off, v79, s32 offset:488
-; GFX11-NEXT: scratch_store_b32 off, v88, s32 offset:484
-; GFX11-NEXT: scratch_store_b32 off, v89, s32 offset:480
-; GFX11-NEXT: scratch_store_b32 off, v90, s32 offset:476
-; GFX11-NEXT: scratch_store_b32 off, v91, s32 offset:472
-; GFX11-NEXT: scratch_store_b32 off, v92, s32 offset:468
-; GFX11-NEXT: scratch_store_b32 off, v93, s32 offset:464
-; GFX11-NEXT: scratch_store_b32 off, v94, s32 offset:460
-; GFX11-NEXT: scratch_store_b32 off, v95, s32 offset:456
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_store_b32 off, v104, s32 offset:452
-; GFX11-NEXT: scratch_store_b32 off, v105, s32 offset:448
-; GFX11-NEXT: scratch_store_b32 off, v106, s32 offset:444
-; GFX11-NEXT: scratch_store_b32 off, v107, s32 offset:440
-; GFX11-NEXT: scratch_store_b32 off, v108, s32 offset:436
-; GFX11-NEXT: scratch_store_b32 off, v109, s32 offset:432
-; GFX11-NEXT: scratch_store_b32 off, v110, s32 offset:428
-; GFX11-NEXT: scratch_store_b32 off, v111, s32 offset:424
-; GFX11-NEXT: scratch_store_b32 off, v120, s32 offset:420
-; GFX11-NEXT: scratch_store_b32 off, v121, s32 offset:416
-; GFX11-NEXT: scratch_store_b32 off, v122, s32 offset:412
-; GFX11-NEXT: scratch_store_b32 off, v123, s32 offset:408
-; GFX11-NEXT: scratch_store_b32 off, v124, s32 offset:404
-; GFX11-NEXT: scratch_store_b32 off, v125, s32 offset:400
-; GFX11-NEXT: scratch_store_b32 off, v126, s32 offset:396
-; GFX11-NEXT: scratch_store_b32 off, v127, s32 offset:392
-; GFX11-NEXT: v_dual_mov_b32 v37, v30 :: v_dual_mov_b32 v54, v24
-; GFX11-NEXT: v_dual_mov_b32 v55, v28 :: v_dual_mov_b32 v52, v26
-; GFX11-NEXT: v_dual_mov_b32 v50, v22 :: v_dual_mov_b32 v53, v20
-; GFX11-NEXT: v_dual_mov_b32 v34, v18 :: v_dual_mov_b32 v51, v16
-; GFX11-NEXT: v_dual_mov_b32 v36, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-NEXT: v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v39, v4
-; GFX11-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v33, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-NEXT: scratch_load_u16 v118, off, s32 offset:380
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-NEXT: scratch_load_u16 v134, off, s32 offset:372
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:364
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:356
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-NEXT: scratch_load_u16 v98, off, s32 offset:348
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:340
-; GFX11-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:332
-; GFX11-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-NEXT: scratch_load_u16 v112, off, s32 offset:324
-; GFX11-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-NEXT: scratch_load_u16 v81, off, s32 offset:316
-; GFX11-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:308
-; GFX11-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:300
-; GFX11-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:292
-; GFX11-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:284
-; GFX11-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:276
-; GFX11-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:268
-; GFX11-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:260
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:252
-; GFX11-NEXT: scratch_load_u16 v94, off, s32 offset:248
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:244
-; GFX11-NEXT: scratch_load_u16 v88, off, s32 offset:240
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:236
-; GFX11-NEXT: scratch_load_u16 v93, off, s32 offset:232
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:228
-; GFX11-NEXT: scratch_load_u16 v91, off, s32 offset:224
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:220
-; GFX11-NEXT: scratch_load_u16 v92, off, s32 offset:216
-; GFX11-NEXT: scratch_load_b32 v150, off, s32 offset:388
-; GFX11-NEXT: scratch_load_u16 v182, off, s32
-; GFX11-NEXT: scratch_load_u16 v40, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v43, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v44, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v45, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v46, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v47, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v58, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v59, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v60, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v61, off, s32 offset:88
-; GFX11-NEXT: scratch_load_u16 v62, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v63, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v72, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v73, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v74, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v75, off, s32 offset:136
-; GFX11-NEXT: scratch_load_u16 v76, off, s32 offset:144
-; GFX11-NEXT: scratch_load_u16 v77, off, s32 offset:152
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v78, off, s32 offset:160
-; GFX11-NEXT: scratch_load_u16 v79, off, s32 offset:168
-; GFX11-NEXT: scratch_load_u16 v89, off, s32 offset:176
-; GFX11-NEXT: scratch_load_u16 v90, off, s32 offset:184
-; GFX11-NEXT: scratch_load_u16 v95, off, s32 offset:192
-; GFX11-NEXT: scratch_load_u16 v104, off, s32 offset:200
-; GFX11-NEXT: scratch_load_u16 v105, off, s32 offset:208
-; GFX11-NEXT: scratch_load_u16 v42, off, s32 offset:212
-; GFX11-NEXT: scratch_load_u16 v180, off, s32 offset:204
-; GFX11-NEXT: scratch_load_u16 v183, off, s32 offset:196
-; GFX11-NEXT: scratch_load_u16 v165, off, s32 offset:188
-; GFX11-NEXT: scratch_load_u16 v179, off, s32 offset:180
-; GFX11-NEXT: scratch_load_u16 v163, off, s32 offset:172
-; GFX11-NEXT: scratch_load_u16 v177, off, s32 offset:164
-; GFX11-NEXT: scratch_load_u16 v145, off, s32 offset:156
-; GFX11-NEXT: scratch_load_u16 v166, off, s32 offset:148
-; GFX11-NEXT: scratch_load_u16 v148, off, s32 offset:140
-; GFX11-NEXT: scratch_load_u16 v151, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v164, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v146, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v162, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v144, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v149, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v119, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v147, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v135, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v133, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v128, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v132, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v161, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v160, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v176, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v167, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v181, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(62)
-; GFX11-NEXT: v_lshlrev_b16 v127, 8, v0
-; GFX11-NEXT: v_lshlrev_b16 v126, 8, v2
-; GFX11-NEXT: v_lshlrev_b16 v124, 8, v4
-; GFX11-NEXT: v_lshlrev_b16 v125, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v120, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v123, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v121, 8, v12
-; GFX11-NEXT: v_lshlrev_b16 v122, 8, v14
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: v_lshlrev_b16 v106, 8, v16
-; GFX11-NEXT: v_lshlrev_b16 v111, 8, v18
-; GFX11-NEXT: v_lshlrev_b16 v109, 8, v20
-; GFX11-NEXT: v_lshlrev_b16 v110, 8, v22
-; GFX11-NEXT: v_lshlrev_b16 v107, 8, v24
-; GFX11-NEXT: v_lshlrev_b16 v108, 8, v26
-; GFX11-NEXT: s_waitcnt vmcnt(61)
-; GFX11-NEXT: v_lshlrev_b16 v88, 8, v88
-; GFX11-NEXT: s_waitcnt vmcnt(59)
-; GFX11-NEXT: v_lshlrev_b16 v93, 8, v93
-; GFX11-NEXT: s_waitcnt vmcnt(57)
-; GFX11-NEXT: v_lshlrev_b16 v91, 8, v91
-; GFX11-NEXT: s_waitcnt vmcnt(55)
-; GFX11-NEXT: v_lshlrev_b16 v92, 8, v92
-; GFX11-NEXT: s_waitcnt vmcnt(54)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v150
-; GFX11-NEXT: s_waitcnt vmcnt(53)
-; GFX11-NEXT: v_lshlrev_b16 v150, 8, v182
-; GFX11-NEXT: s_waitcnt vmcnt(52)
-; GFX11-NEXT: v_lshlrev_b16 v41, 8, v40
-; GFX11-NEXT: s_waitcnt vmcnt(51)
-; GFX11-NEXT: v_lshlrev_b16 v40, 8, v43
-; GFX11-NEXT: s_waitcnt vmcnt(50)
-; GFX11-NEXT: v_lshlrev_b16 v43, 8, v44
-; GFX11-NEXT: s_waitcnt vmcnt(49)
-; GFX11-NEXT: v_lshlrev_b16 v182, 8, v45
-; GFX11-NEXT: s_waitcnt vmcnt(48)
-; GFX11-NEXT: v_lshlrev_b16 v46, 8, v46
-; GFX11-NEXT: s_waitcnt vmcnt(47)
-; GFX11-NEXT: v_lshlrev_b16 v45, 8, v47
-; GFX11-NEXT: s_waitcnt vmcnt(46)
-; GFX11-NEXT: v_lshlrev_b16 v57, 8, v56
-; GFX11-NEXT: s_waitcnt vmcnt(45)
-; GFX11-NEXT: v_lshlrev_b16 v56, 8, v58
-; GFX11-NEXT: s_waitcnt vmcnt(44)
-; GFX11-NEXT: v_lshlrev_b16 v58, 8, v59
-; GFX11-NEXT: s_waitcnt vmcnt(43)
-; GFX11-NEXT: v_lshlrev_b16 v44, 8, v60
-; GFX11-NEXT: s_waitcnt vmcnt(42)
-; GFX11-NEXT: v_lshlrev_b16 v60, 8, v61
-; GFX11-NEXT: s_waitcnt vmcnt(41)
-; GFX11-NEXT: v_lshlrev_b16 v59, 8, v62
-; GFX11-NEXT: s_waitcnt vmcnt(40)
-; GFX11-NEXT: v_lshlrev_b16 v62, 8, v63
-; GFX11-NEXT: s_waitcnt vmcnt(39)
-; GFX11-NEXT: v_lshlrev_b16 v47, 8, v72
-; GFX11-NEXT: s_waitcnt vmcnt(38)
-; GFX11-NEXT: v_lshlrev_b16 v72, 8, v73
-; GFX11-NEXT: s_waitcnt vmcnt(37)
-; GFX11-NEXT: v_lshlrev_b16 v63, 8, v74
-; GFX11-NEXT: s_waitcnt vmcnt(36)
-; GFX11-NEXT: v_lshlrev_b16 v74, 8, v75
-; GFX11-NEXT: s_waitcnt vmcnt(35)
-; GFX11-NEXT: v_lshlrev_b16 v73, 8, v76
-; GFX11-NEXT: s_waitcnt vmcnt(34)
-; GFX11-NEXT: v_lshlrev_b16 v75, 8, v77
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v61, 8, v78
-; GFX11-NEXT: s_waitcnt vmcnt(32)
-; GFX11-NEXT: v_lshlrev_b16 v78, 8, v79
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v77, 8, v89
-; GFX11-NEXT: s_waitcnt vmcnt(30)
-; GFX11-NEXT: v_lshlrev_b16 v79, 8, v90
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v76, 8, v95
-; GFX11-NEXT: s_waitcnt vmcnt(28)
-; GFX11-NEXT: v_lshlrev_b16 v90, 8, v104
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v89, 8, v105
-; GFX11-NEXT: v_lshlrev_b16 v104, 8, v94
-; GFX11-NEXT: v_lshlrev_b16 v95, 8, v31
-; GFX11-NEXT: v_lshlrev_b16 v105, 8, v30
-; GFX11-NEXT: v_lshlrev_b16 v94, 8, v28
-; GFX11-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v48
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v70
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v71
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v84
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v82
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v51
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v34
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v52
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v66
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v128
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v113
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v132
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v100
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v161
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v160
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v176
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v167
-; GFX11-NEXT: v_perm_b32 v2, v2, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v5, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v7, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v9, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v11, v10, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v102
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v114
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v96
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v133
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v117
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v135
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v130
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v181
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v150
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v41
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v40
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v43
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v182
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v46
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v45
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v57
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v56
-; GFX11-NEXT: v_perm_b32 v7, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v12, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v14, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v16, v15, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v147
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v119
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v144
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v162
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v146
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v164
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v151
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v148
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v58
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v44
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v60
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v59
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v62
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v47
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v72
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v63
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v74
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v73
-; GFX11-NEXT: v_perm_b32 v12, v13, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v15, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v17, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v19, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v21, v20, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v166
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v145
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v163
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v179
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v165
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v183
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v180
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v42
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v65
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v75
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v61
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v78
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v77
-; GFX11-NEXT: v_or_b32_e32 v21, v21, v79
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v76
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v90
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v89
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v92
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v91
-; GFX11-NEXT: v_perm_b32 v17, v18, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v20, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v22, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v24, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v26, v25, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v101
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v22, v22, v93
-; GFX11-NEXT: v_or_b32_e32 v23, v23, v88
-; GFX11-NEXT: v_or_b32_e32 v24, v24, v104
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v95
-; GFX11-NEXT: v_or_b32_e32 v26, v26, v105
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v94
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v108
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v107
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v110
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v109
-; GFX11-NEXT: v_perm_b32 v22, v23, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v25, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v27, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v29, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v31, v30, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v103
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v112
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v99
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v129
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v98
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v131
-; GFX11-NEXT: v_and_b32_e32 v34, 0xff, v116
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v134
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-NEXT: v_or_b32_e32 v27, v27, v111
-; GFX11-NEXT: v_or_b32_e32 v28, v28, v106
-; GFX11-NEXT: v_or_b32_e32 v29, v29, v122
-; GFX11-NEXT: v_or_b32_e32 v30, v30, v121
-; GFX11-NEXT: v_or_b32_e32 v31, v31, v123
-; GFX11-NEXT: v_or_b32_e32 v32, v32, v120
-; GFX11-NEXT: v_or_b32_e32 v33, v33, v125
-; GFX11-NEXT: v_or_b32_e32 v34, v34, v124
-; GFX11-NEXT: v_or_b32_e32 v35, v35, v126
-; GFX11-NEXT: v_or_b32_e32 v36, v36, v127
-; GFX11-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v30, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v32, v31, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v34, v33, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v36, v35, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr78
-; GFX11-NEXT: ; implicit-def: $vgpr77
-; GFX11-NEXT: ; implicit-def: $vgpr79
-; GFX11-NEXT: ; implicit-def: $vgpr76
-; GFX11-NEXT: ; implicit-def: $vgpr90
-; GFX11-NEXT: ; implicit-def: $vgpr89
-; GFX11-NEXT: ; implicit-def: $vgpr92
-; GFX11-NEXT: ; implicit-def: $vgpr91
-; GFX11-NEXT: ; implicit-def: $vgpr93
-; GFX11-NEXT: ; implicit-def: $vgpr88
-; GFX11-NEXT: ; implicit-def: $vgpr104
-; GFX11-NEXT: ; implicit-def: $vgpr95
-; GFX11-NEXT: ; implicit-def: $vgpr105
-; GFX11-NEXT: ; implicit-def: $vgpr94
-; GFX11-NEXT: ; implicit-def: $vgpr108
-; GFX11-NEXT: ; implicit-def: $vgpr107
-; GFX11-NEXT: ; implicit-def: $vgpr110
-; GFX11-NEXT: ; implicit-def: $vgpr109
-; GFX11-NEXT: ; implicit-def: $vgpr111
-; GFX11-NEXT: ; implicit-def: $vgpr106
-; GFX11-NEXT: ; implicit-def: $vgpr122
-; GFX11-NEXT: ; implicit-def: $vgpr121
-; GFX11-NEXT: ; implicit-def: $vgpr123
-; GFX11-NEXT: ; implicit-def: $vgpr120
-; GFX11-NEXT: ; implicit-def: $vgpr125
-; GFX11-NEXT: ; implicit-def: $vgpr124
-; GFX11-NEXT: ; implicit-def: $vgpr126
-; GFX11-NEXT: ; implicit-def: $vgpr127
-; GFX11-NEXT: .LBB48_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v134, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v118, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v131, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v116, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v129, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v35, v35, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v126, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v127, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v125, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v124, v3
-; GFX11-NEXT: v_add_nc_u16 v33, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v31, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v98, 3
-; GFX11-NEXT: v_add_nc_u16 v116, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v30, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v98, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v1, v112, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_add_nc_u16 v3, v99, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v103, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v120, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v122, v1
-; GFX11-NEXT: v_add_nc_u16 v29, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v121, v3
-; GFX11-NEXT: v_add_nc_u16 v99, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v111, v4
-; GFX11-NEXT: v_add_nc_u16 v28, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v81, 3
-; GFX11-NEXT: v_add_nc_u16 v81, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v101, 3
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v86, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v97, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v83, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v106, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v110, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v109, v0
-; GFX11-NEXT: v_add_nc_u16 v83, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v108, v3
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v107, v4
-; GFX11-NEXT: v_add_nc_u16 v86, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v80, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v69, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v105, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v94, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v104, v2
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v95, v3
-; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v93, v4
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(26)
-; GFX11-NEXT: v_add_nc_u16 v0, v42, 3
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v65, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(24)
-; GFX11-NEXT: v_add_nc_u16 v3, v183, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v180, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v88, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v92, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v91, v1
-; GFX11-NEXT: v_add_nc_u16 v65, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v90, v3
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v89, v4
-; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_add_nc_u16 v1, v179, 3
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v165, 3
-; GFX11-NEXT: v_add_nc_u16 v80, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_add_nc_u16 v0, v177, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v163, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_add_nc_u16 v4, v166, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v79, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v76, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v78, v0
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v77, v3
-; GFX11-NEXT: v_add_nc_u16 v85, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v75, v4
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v145, 3
-; GFX11-NEXT: v_add_nc_u16 v97, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_add_nc_u16 v1, v151, 3
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v148, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_add_nc_u16 v3, v178, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v164, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v61, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v74, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v73, v2
-; GFX11-NEXT: v_add_nc_u16 v101, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v72, v3
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v63, v4
-; GFX11-NEXT: v_add_nc_u16 v103, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_add_nc_u16 v2, v162, 3
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v146, 3
-; GFX11-NEXT: v_add_nc_u16 v112, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(10)
-; GFX11-NEXT: v_add_nc_u16 v1, v149, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v144, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_add_nc_u16 v4, v147, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v62, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v47, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v60, v1
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v59, v3
-; GFX11-NEXT: v_add_nc_u16 v118, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v58, v4
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v119, 3
-; GFX11-NEXT: v_add_nc_u16 v119, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_add_nc_u16 v2, v135, 3
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v130, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_add_nc_u16 v3, v133, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v117, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v44, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v57, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v56, v0
-; GFX11-NEXT: v_add_nc_u16 v117, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v46, v3
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v45, v4
-; GFX11-NEXT: v_add_nc_u16 v129, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_add_nc_u16 v0, v114, 3
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v96, 3
-; GFX11-NEXT: v_add_nc_u16 v96, 0x300, v2
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v2, v102, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v87, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v55, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v43, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v182, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v41, v2
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v40, v3
-; GFX11-NEXT: v_add_nc_u16 v55, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v181, v4
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v37, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v52, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v53, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v50, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v150, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v176, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v167, v1
-; GFX11-NEXT: v_add_nc_u16 v50, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v161, v3
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v160, v4
-; GFX11-NEXT: v_add_nc_u16 v52, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v49, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v48, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v36, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v132, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v128, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v115, v0
-; GFX11-NEXT: v_add_nc_u16 v34, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v100, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v113, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v39, 3
-; GFX11-NEXT: v_add_nc_u16 v36, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v32, v32, 3
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-NEXT: v_or_b32_e32 v35, v71, v35
-; GFX11-NEXT: v_or_b32_e32 v33, v82, v33
-; GFX11-NEXT: v_or_b32_e32 v0, v84, v0
-; GFX11-NEXT: v_or_b32_e32 v36, v70, v36
-; GFX11-NEXT: v_or_b32_e32 v32, v66, v32
-; GFX11-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-NEXT: v_add_nc_u16 v38, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v36
-; GFX11-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-NEXT: v_add_nc_u16 v36, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v39, 0x300, v1
-; GFX11-NEXT: v_perm_b32 v1, v33, v38, 0x5040100
-; GFX11-NEXT: v_perm_b32 v0, v35, v0, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v32, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v36, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v39, v34, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v52, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v50, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v96, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v129, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v12, v117, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v119, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v118, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v112, v15, 0x5040100
-; GFX11-NEXT: v_perm_b32 v16, v103, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v17, v101, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v18, v97, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v19, v85, v19, 0x5040100
-; GFX11-NEXT: v_perm_b32 v20, v80, v20, 0x5040100
-; GFX11-NEXT: v_perm_b32 v21, v69, v21, 0x5040100
-; GFX11-NEXT: v_perm_b32 v22, v65, v22, 0x5040100
-; GFX11-NEXT: v_perm_b32 v23, v64, v23, 0x5040100
-; GFX11-NEXT: v_perm_b32 v24, v68, v24, 0x5040100
-; GFX11-NEXT: v_perm_b32 v25, v67, v25, 0x5040100
-; GFX11-NEXT: v_perm_b32 v26, v86, v26, 0x5040100
-; GFX11-NEXT: v_perm_b32 v27, v83, v27, 0x5040100
-; GFX11-NEXT: v_perm_b32 v28, v81, v28, 0x5040100
-; GFX11-NEXT: v_perm_b32 v29, v99, v29, 0x5040100
-; GFX11-NEXT: v_perm_b32 v30, v98, v30, 0x5040100
-; GFX11-NEXT: v_perm_b32 v31, v116, v31, 0x5040100
-; GFX11-NEXT: .LBB48_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_b32 v127, off, s32 offset:392
-; GFX11-NEXT: scratch_load_b32 v126, off, s32 offset:396
-; GFX11-NEXT: scratch_load_b32 v125, off, s32 offset:400
-; GFX11-NEXT: scratch_load_b32 v124, off, s32 offset:404
-; GFX11-NEXT: scratch_load_b32 v123, off, s32 offset:408
-; GFX11-NEXT: scratch_load_b32 v122, off, s32 offset:412
-; GFX11-NEXT: scratch_load_b32 v121, off, s32 offset:416
-; GFX11-NEXT: scratch_load_b32 v120, off, s32 offset:420
-; GFX11-NEXT: scratch_load_b32 v111, off, s32 offset:424
-; GFX11-NEXT: scratch_load_b32 v110, off, s32 offset:428
-; GFX11-NEXT: scratch_load_b32 v109, off, s32 offset:432
-; GFX11-NEXT: scratch_load_b32 v108, off, s32 offset:436
-; GFX11-NEXT: scratch_load_b32 v107, off, s32 offset:440
-; GFX11-NEXT: scratch_load_b32 v106, off, s32 offset:444
-; GFX11-NEXT: scratch_load_b32 v105, off, s32 offset:448
-; GFX11-NEXT: scratch_load_b32 v104, off, s32 offset:452
-; GFX11-NEXT: scratch_load_b32 v95, off, s32 offset:456
-; GFX11-NEXT: scratch_load_b32 v94, off, s32 offset:460
-; GFX11-NEXT: scratch_load_b32 v93, off, s32 offset:464
-; GFX11-NEXT: scratch_load_b32 v92, off, s32 offset:468
-; GFX11-NEXT: scratch_load_b32 v91, off, s32 offset:472
-; GFX11-NEXT: scratch_load_b32 v90, off, s32 offset:476
-; GFX11-NEXT: scratch_load_b32 v89, off, s32 offset:480
-; GFX11-NEXT: scratch_load_b32 v88, off, s32 offset:484
-; GFX11-NEXT: scratch_load_b32 v79, off, s32 offset:488
-; GFX11-NEXT: scratch_load_b32 v78, off, s32 offset:492
-; GFX11-NEXT: scratch_load_b32 v77, off, s32 offset:496
-; GFX11-NEXT: scratch_load_b32 v76, off, s32 offset:500
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:504
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:508
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:512
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:516
-; GFX11-NEXT: s_clause 0xf
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:520
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:524
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:528
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:532
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:536
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:540
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:544
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:548
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:552
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:556
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:560
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:564
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:568
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:572
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:576
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:580
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <128 x i8> %a, splat (i8 3)
- %a2 = bitcast <128 x i8> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <128 x i8> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v128i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:128
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:100
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:92
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:84
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:76
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v15
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:24
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v14
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v18
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(10) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v27
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v7
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v12
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(8) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v49
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v48
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v39
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v11
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v37
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; kill: killed $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; kill: killed $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; kill: killed $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; kill: killed $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; kill: killed $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; kill: killed $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; kill: killed $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; kill: killed $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; kill: killed $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GCN-NEXT: v_or_b32_e32 v54, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v50, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v8
-; GCN-NEXT: v_or_b32_e32 v51, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v52, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v12
-; GCN-NEXT: v_or_b32_e32 v53, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v55, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; GCN-NEXT: v_or_b32_e32 v41, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v42, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: v_or_b32_e32 v43, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v24
-; GCN-NEXT: v_or_b32_e32 v45, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v46, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; GCN-NEXT: v_or_b32_e32 v47, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v56, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v7
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v60, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_or_b32_e32 v61, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v62, v1, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v49
-; GCN-NEXT: v_or_b32_e32 v63, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v3, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v48
-; GCN-NEXT: v_or_b32_e32 v6, v5, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v10, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v39
-; GCN-NEXT: v_or_b32_e32 v14, v13, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v15, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: v_or_b32_e32 v22, v17, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v26, v19, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v30, v21, v27
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v32, v23, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v33, v33, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v4, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v8, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v12, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v16, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v20, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v24, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v28, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v31, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v7, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v9, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v49, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v48, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v39, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v11, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v38, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v1, v37, 8, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v5, v54, v40, 24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v40, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v40, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v50, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v50, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v50, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v52, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v52, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v52, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v55, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v55, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v55, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v44, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v44, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v44, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v46, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v46, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v46, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; kill: killed $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: .LBB49_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v29, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v4, v25, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v34, v6
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v27, v7
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v5
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v36
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v33, v8
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v33, v9
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v33, v10
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v33, v11
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v33, v12
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v33, v14
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v14, v16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v14, v17
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v14, v15
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v35, v14, v18
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v14, v5
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v14, v13
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v14, v19
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v14, v20
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v14, v21
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v36, v14, v22
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v14, v23
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v14, v24
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v14, v25
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v37, v14, v26
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v14, v27
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v14, v28
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v14, v29
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v38, v14, v30
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v14, v31
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v39, v14, v32
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v14, v1
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v14, v3
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v33, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v30, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v34
-; GCN-NEXT: v_add_i32_e32 v62, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v63, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v60, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v61, vcc, s6, v35
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v59, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v57, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v46, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v47, vcc, s6, v36
-; GCN-NEXT: v_add_i32_e32 v44, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v45, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v42, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v43, vcc, s6, v37
-; GCN-NEXT: v_add_i32_e32 v55, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v41, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v53, vcc, s6, v38
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v31
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v39
-; GCN-NEXT: v_add_i32_e32 v40, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v3
-; GCN-NEXT: v_alignbit_b32 v5, v54, v40, 24
-; GCN-NEXT: v_alignbit_b32 v9, v54, v40, 16
-; GCN-NEXT: v_alignbit_b32 v4, v54, v40, 8
-; GCN-NEXT: v_alignbit_b32 v1, v51, v50, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v50, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v8, v51, v50, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v52, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v52, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v11, v53, v52, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v55, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v41, v55, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v12, v41, v55, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v43, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v43, v42, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v44, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v45, v44, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v24, v45, v44, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v46, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v47, v46, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v28, v47, v46, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v57, v56, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v59, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v61, v60, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v63, v62, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v2, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v14, v10, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v22, v18, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v30, v26, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v51
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v53
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v41
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v43
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v45
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB49_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v40
-; GCN-NEXT: s_waitcnt vmcnt(6)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v4
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v54
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v5, v7, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v1, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v3, v5
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v50
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v8
-; GCN-NEXT: v_or_b32_e32 v29, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v31
-; GCN-NEXT: v_or_b32_e32 v31, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v11
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v53
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v12
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v41
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v20
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v43
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v24
-; GCN-NEXT: v_or_b32_e32 v9, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v45
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v28
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v47
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v56
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v15, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v57
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v16, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v58
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v17, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v60
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v20, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v61
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v21, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v62
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v23, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v63
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v24, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v10
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v10, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v14
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v14, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v18, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v22
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v22, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v26
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v25, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v30
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v26, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v27, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v33
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; GCN-NEXT: v_or_b32_e32 v28, v1, v2
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v30, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v33, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v32, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v34, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v35, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v39
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v36, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v37, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v48
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v38, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v39, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v49
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v48, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v49, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v50, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v51, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v52, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v53, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v54, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v55, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v40, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v41, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v42, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v43, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v56, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v57, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v58, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v59, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v60, v2, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v4, v1, v30
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v5, v1, v33
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v63, v2, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 12, v0
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v61, v3, v34
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v35
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v6, v6, v36
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v37
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 28, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v7, v38
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 32, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v8, v7, v39
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v9, v7, v48
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v7, v49
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v12, v7, v50
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v13, v7, v51
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v15, v7, v52
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v16, v7, v53
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v17, v7, v54
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v7, v55
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v20, v7, v40
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v21, v7, v41
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v23, v7, v42
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v51
-; GCN-NEXT: v_or_b32_e32 v44, v51, v44
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v10, v45
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v14, v46
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v18, v18, v47
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v22, v22, v56
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_or_b32_e32 v25, v25, v57
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v26, v26, v58
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x70, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v27, v27, v59
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v28, v28, v60
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v63, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v6, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v128i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v36, off, s[0:3], s32
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v13
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v11
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v9
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v5
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v3
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v1
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v29
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v28
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v27
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v26
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v25
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v24
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v59, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v57, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v30
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
-; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v21
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v20
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v18
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v37
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v36
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v17
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; kill: killed $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; kill: killed $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; kill: killed $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; kill: killed $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; kill: killed $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; kill: killed $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; kill: killed $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v16
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v14
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v13
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v12
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v11
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v10
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v10
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v9
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v8
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v8
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v7
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v6
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[31:32], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v31, v7
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v10
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v11
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v12
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v13
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v14
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v16
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v9, v8
-; VI-NEXT: v_lshrrev_b64 v[7:8], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v7, v5
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v7, v6
-; VI-NEXT: v_lshrrev_b64 v[5:6], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v5, 24, v4
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v3
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v5, v3
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v2
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v3, 8, v2
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v3, 8, v1
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v3, v1
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 24, v37
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v36
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v36
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v37
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v3, v2
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[36:37]
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v5, v4
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 24, v30
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v30
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 24, v28
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v28
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v27
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v27
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v28
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v26
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v26
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v24
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v23
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v23
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v24
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v21
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v21
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v22
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v19
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v20
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v17
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v18
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[23:24]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[21:22]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v1, v46
-; VI-NEXT: v_lshrrev_b64 v[45:46], 24, v[17:18]
-; VI-NEXT: v_mov_b32_e32 v32, v15
-; VI-NEXT: v_lshrrev_b32_e32 v35, 24, v26
-; VI-NEXT: v_lshrrev_b32_e32 v39, 24, v24
-; VI-NEXT: v_lshrrev_b32_e32 v58, 24, v22
-; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v20
-; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v20
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v18
-; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v18
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v17
-; VI-NEXT: v_mov_b32_e32 v46, v1
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: .LBB49_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v31, 3
-; VI-NEXT: v_add_u16_sdwa v51, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v32, 3, v18
-; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v51
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v18, v32, v18
-; VI-NEXT: v_add_u16_e32 v32, 3, v17
-; VI-NEXT: v_add_u16_sdwa v17, v17, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v17, v32, v17
-; VI-NEXT: v_add_u16_e32 v32, 3, v20
-; VI-NEXT: v_add_u16_sdwa v20, v20, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v20, v32, v20
-; VI-NEXT: v_add_u16_e32 v32, 3, v19
-; VI-NEXT: v_add_u16_sdwa v19, v19, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; VI-NEXT: v_add_u16_sdwa v48, v22, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v19, v32, v19
-; VI-NEXT: v_add_u16_e32 v32, 3, v22
-; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v48
-; VI-NEXT: v_add_u16_sdwa v53, v21, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v22, v32, v22
-; VI-NEXT: v_add_u16_e32 v32, 3, v21
-; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v53
-; VI-NEXT: v_add_u16_sdwa v61, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v21, v32, v21
-; VI-NEXT: v_add_u16_e32 v32, 3, v24
-; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v61
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v24, v32, v24
-; VI-NEXT: v_add_u16_e32 v32, 3, v23
-; VI-NEXT: v_add_u16_sdwa v23, v23, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; VI-NEXT: v_add_u16_sdwa v58, v26, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v23, v32, v23
-; VI-NEXT: v_add_u16_e32 v32, 3, v26
-; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v58
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v26, v32, v26
-; VI-NEXT: v_add_u16_e32 v32, 3, v25
-; VI-NEXT: v_add_u16_sdwa v25, v25, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; VI-NEXT: v_add_u16_sdwa v39, v28, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v25, v32, v25
-; VI-NEXT: v_add_u16_e32 v32, 3, v28
-; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v28, v32, v28
-; VI-NEXT: v_add_u16_e32 v32, 3, v27
-; VI-NEXT: v_add_u16_sdwa v27, v27, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; VI-NEXT: v_add_u16_sdwa v35, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_add_u16_e32 v33, 3, v30
-; VI-NEXT: v_add_u16_e32 v34, 3, v29
-; VI-NEXT: v_add_u16_sdwa v32, v29, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v35
-; VI-NEXT: v_add_u16_sdwa v52, v37, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v30, v33, v29
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v32
-; VI-NEXT: v_add_u16_e32 v33, 3, v37
-; VI-NEXT: v_add_u16_sdwa v50, v36, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v29, v34, v29
-; VI-NEXT: v_add_u16_e32 v34, 3, v36
-; VI-NEXT: v_or_b32_e32 v37, v33, v32
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v50
-; VI-NEXT: v_add_u16_sdwa v57, v2, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v36, v34, v32
-; VI-NEXT: v_add_u16_e32 v33, 3, v2
-; VI-NEXT: v_add_u16_e32 v34, 3, v1
-; VI-NEXT: v_add_u16_sdwa v32, v1, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v57
-; VI-NEXT: v_or_b32_e32 v2, v33, v1
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v32
-; VI-NEXT: v_add_u16_sdwa v56, v4, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v1, v34, v1
-; VI-NEXT: v_add_u16_e32 v33, 3, v4
-; VI-NEXT: v_add_u16_e32 v34, 3, v3
-; VI-NEXT: v_add_u16_sdwa v32, v3, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v56
-; VI-NEXT: v_or_b32_e32 v4, v33, v3
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
-; VI-NEXT: v_add_u16_sdwa v47, v6, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v3, v34, v3
-; VI-NEXT: v_add_u16_e32 v33, 3, v6
-; VI-NEXT: v_add_u16_e32 v34, 3, v5
-; VI-NEXT: v_add_u16_sdwa v32, v5, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v47
-; VI-NEXT: v_or_b32_e32 v6, v33, v5
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v32
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v5, v34, v5
-; VI-NEXT: v_add_u16_sdwa v34, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: v_add_u16_e32 v38, 3, v8
-; VI-NEXT: v_add_u16_e32 v33, 3, v7
-; VI-NEXT: v_add_u16_sdwa v32, v7, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v34
-; VI-NEXT: v_or_b32_e32 v8, v38, v7
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v32
-; VI-NEXT: v_add_u16_sdwa v59, v10, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v7, v33, v7
-; VI-NEXT: v_add_u16_e32 v33, 3, v10
-; VI-NEXT: v_add_u16_e32 v38, 3, v9
-; VI-NEXT: v_add_u16_sdwa v32, v9, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v59
-; VI-NEXT: v_or_b32_e32 v10, v33, v9
-; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v32
-; VI-NEXT: v_add_u16_sdwa v63, v12, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v9, v38, v9
-; VI-NEXT: v_add_u16_e32 v33, 3, v12
-; VI-NEXT: v_add_u16_e32 v38, 3, v11
-; VI-NEXT: v_add_u16_sdwa v32, v11, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v63
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v12, v33, v11
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v32
-; VI-NEXT: v_add_u16_sdwa v33, v14, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v11, v38, v11
-; VI-NEXT: v_add_u16_e32 v38, 3, v14
-; VI-NEXT: v_add_u16_e32 v49, 3, v13
-; VI-NEXT: v_add_u16_sdwa v32, v13, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v33
-; VI-NEXT: v_add_u16_sdwa v60, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v38, v13
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v32
-; VI-NEXT: v_add_u16_sdwa v31, v15, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_add_u16_e32 v32, 3, v15
-; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v60
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v16, v16, v15
-; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v31
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v15, v32, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v16
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_or_b32_e32 v13, v49, v13
-; VI-NEXT: v_lshrrev_b32_e32 v15, 8, v14
-; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v15, 8, v13
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[11:12]
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v11, 8, v10
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v11, 8, v9
-; VI-NEXT: v_lshrrev_b64 v[9:10], 24, v[9:10]
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v8
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; VI-NEXT: v_lshrrev_b64 v[7:8], 24, v[7:8]
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v6
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v5
-; VI-NEXT: v_lshrrev_b64 v[5:6], 24, v[5:6]
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v3
-; VI-NEXT: v_lshrrev_b64 v[40:41], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v3, 8, v2
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v3, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[43:44], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v36
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[36:37]
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v29
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[29:30]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v28
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v27
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[27:28]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v26
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v25
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[1:2], 24, v[25:26]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v24
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v23
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v21
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v60, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v33, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v63, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v59, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v34, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v47, 8, 8
-; VI-NEXT: v_lshrrev_b64 v[44:45], 24, v[19:20]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v56, 8, 8
-; VI-NEXT: v_lshrrev_b64 v[45:46], 24, v[17:18]
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v57, 8, 8
-; VI-NEXT: v_mov_b32_e32 v46, v35
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v52, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v46, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v39, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v49, v53
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v52, v51
-; VI-NEXT: v_bfe_u32 v31, v51, 8, 8
-; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: v_lshrrev_b64 v[54:55], 24, v[23:24]
-; VI-NEXT: v_lshrrev_b64 v[41:42], 24, v[21:22]
-; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v20
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v18
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v17
-; VI-NEXT: v_bfe_u32 v35, v58, 8, 8
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v39, v61, 8, 8
-; VI-NEXT: v_bfe_u32 v58, v48, 8, 8
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_bfe_u32 v61, v53, 8, 8
-; VI-NEXT: .LBB49_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v40
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v56, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v34, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v59, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v63, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v60, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v55
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v45
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v31
-; VI-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v41
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v58
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v54
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v39
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v35
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v46, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x74, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x78, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v128i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; kill: killed $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; kill: killed $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(45)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(46)
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[31:32]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[29:30]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[27:28]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[25:26]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[23:24]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[21:22]
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[19:20]
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 24, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v17
-; GFX9-NEXT: v_lshrrev_b64 v[42:43], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; GFX9-NEXT: .LBB49_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[29:30]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[27:28]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[25:26]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[23:24]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[33:34], 24, v[21:22]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v16
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v15
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v13
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v12
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v11
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v10
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v10
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v8
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v8
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v7
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v6
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v5
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v4
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v3
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v2
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v31
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v30
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v29
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v28
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v27
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v26
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v25
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v24
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v23
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v22
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
-; GFX9-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v22
-; GFX9-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v21
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[40:41], 24, v[5:6]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v21
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[41:42], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[53:54], 24, v[19:20]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 24, v20
-; GFX9-NEXT: v_lshrrev_b64 v[42:43], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[54:55], 24, v[17:18]
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v32
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 8, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 24, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v17
-; GFX9-NEXT: .LBB49_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v63
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v61
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v38
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v36
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v35
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v34
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v17, v17, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v49
-; GFX9-NEXT: v_or_b32_sdwa v18, v18, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v42
-; GFX9-NEXT: v_or_b32_sdwa v34, v58, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v57, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v40
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v51
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v50
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v53
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v33
-; GFX9-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v23, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v25, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v29, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v56
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v128i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: ; implicit-def: $vgpr74
-; GFX11-NEXT: ; implicit-def: $vgpr72
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr63
-; GFX11-NEXT: ; implicit-def: $vgpr62
-; GFX11-NEXT: ; implicit-def: $vgpr60
-; GFX11-NEXT: ; implicit-def: $vgpr57
-; GFX11-NEXT: ; implicit-def: $vgpr47
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr45
-; GFX11-NEXT: ; implicit-def: $vgpr43
-; GFX11-NEXT: ; implicit-def: $vgpr42
-; GFX11-NEXT: ; implicit-def: $vgpr183
-; GFX11-NEXT: ; implicit-def: $vgpr181
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr179
-; GFX11-NEXT: ; implicit-def: $vgpr177
-; GFX11-NEXT: ; implicit-def: $vgpr167
-; GFX11-NEXT: ; implicit-def: $vgpr165
-; GFX11-NEXT: ; implicit-def: $vgpr164
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr161
-; GFX11-NEXT: ; implicit-def: $vgpr151
-; GFX11-NEXT: ; implicit-def: $vgpr150
-; GFX11-NEXT: ; implicit-def: $vgpr147
-; GFX11-NEXT: ; implicit-def: $vgpr145
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr144
-; GFX11-NEXT: ; implicit-def: $vgpr133
-; GFX11-NEXT: ; implicit-def: $vgpr131
-; GFX11-NEXT: ; implicit-def: $vgpr129
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr75
-; GFX11-NEXT: ; implicit-def: $vgpr73
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr61
-; GFX11-NEXT: ; implicit-def: $vgpr59
-; GFX11-NEXT: ; implicit-def: $vgpr58
-; GFX11-NEXT: ; implicit-def: $vgpr56
-; GFX11-NEXT: ; implicit-def: $vgpr46
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr44
-; GFX11-NEXT: ; implicit-def: $vgpr41
-; GFX11-NEXT: ; implicit-def: $vgpr40
-; GFX11-NEXT: ; implicit-def: $vgpr182
-; GFX11-NEXT: ; implicit-def: $vgpr180
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr178
-; GFX11-NEXT: ; implicit-def: $vgpr176
-; GFX11-NEXT: ; implicit-def: $vgpr166
-; GFX11-NEXT: ; implicit-def: $vgpr163
-; GFX11-NEXT: ; implicit-def: $vgpr162
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr160
-; GFX11-NEXT: ; implicit-def: $vgpr149
-; GFX11-NEXT: ; implicit-def: $vgpr148
-; GFX11-NEXT: ; implicit-def: $vgpr146
-; GFX11-NEXT: ; implicit-def: $vgpr135
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr134
-; GFX11-NEXT: ; implicit-def: $vgpr132
-; GFX11-NEXT: ; implicit-def: $vgpr130
-; GFX11-NEXT: ; implicit-def: $vgpr128
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB49_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-NEXT: .LBB49_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB49_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-NEXT: .LBB49_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v74
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v60
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v67, 0xff, v47
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v63
-; GFX11-NEXT: v_or_b32_e32 v54, v67, v54
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v42
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v62
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v55, v55, v64
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v57
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v64
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v45
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v43
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v51
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v54, v64, v67
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v181
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v55
-; GFX11-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v177
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v167
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v54
-; GFX11-NEXT: v_or_b32_e32 v53, v55, v64
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v165
-; GFX11-NEXT: v_and_b32_e32 v55, 0xff, v164
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v54
-; GFX11-NEXT: v_or_b32_e32 v52, v55, v52
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v161
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v151
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v150
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v53
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_or_b32_e32 v52, v54, v55
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v145
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v147
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v144
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: v_or_b32_e32 v51, v54, v51
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v52
-; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v51, 0xff, v133
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v131
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v129
-; GFX11-NEXT: v_and_b32_e32 v54, 0xff, v119
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v117
-; GFX11-NEXT: v_and_b32_e32 v64, 0xff, v116
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v115
-; GFX11-NEXT: v_or_b32_e32 v51, v51, v52
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v53
-; GFX11-NEXT: v_or_b32_e32 v50, v54, v50
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v55
-; GFX11-NEXT: v_or_b32_e32 v52, v64, v67
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-NEXT: v_or_b32_e32 v2, v10, v51
-; GFX11-NEXT: v_or_b32_e32 v3, v11, v50
-; GFX11-NEXT: v_or_b32_e32 v4, v12, v52
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v6, 8, v102
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v100
-; GFX11-NEXT: v_lshlrev_b16 v8, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v10, 8, v98
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v97
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v96
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v86
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v82
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v68
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v75
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v73
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v66
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v39, v48
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v61
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v59
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v58
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v56
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v46
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v65
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v44
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v41
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v40
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v182
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v180
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v178
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v176
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v166
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v163
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v162
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v160
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v148
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v146
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v135
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v134
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v132
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v130
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v114
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v113
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v112
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v103
-; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v101
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v99
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v87
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-NEXT: v_lshlrev_b16 v28, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v81
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v71
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v69
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-NEXT: s_clause 0x5
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-NEXT: s_clause 0x13
-; GFX11-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <128 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <128 x i8>
- br label %end
-
-end:
- %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <128 x i8> %phi
-}
-
-define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:32
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:120
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v42
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v44
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v46
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v47
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v56
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:112
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v57
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v58
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v1
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v61
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:132
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v62
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v1
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v2
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; kill: killed $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: v_mov_b32_e32 v5, v4
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v7
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v9
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v35
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v36
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v37
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v38
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v39
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v48
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v49
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v50
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v51
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v53
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v54
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v55
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v40
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v41
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v43
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v44
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v46
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v47
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v56
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v57
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v58
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v59
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v61
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v62
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v63
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: .LBB50_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v60
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v58
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v56
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v46
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v42
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v42, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v43, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v47, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v56, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v60, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v61, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v62, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v63, 0xffff0000, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v31
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v35
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v39, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v48, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v49, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v50, 0x40c00000, v53
-; GCN-NEXT: v_add_f32_e32 v51, 0x40c00000, v54
-; GCN-NEXT: v_add_f32_e32 v52, 0x40c00000, v55
-; GCN-NEXT: v_add_f32_e32 v53, 0x40c00000, v40
-; GCN-NEXT: v_add_f32_e32 v54, 0x40c00000, v41
-; GCN-NEXT: v_add_f32_e32 v55, 0x40c00000, v42
-; GCN-NEXT: v_add_f32_e32 v40, 0x40c00000, v43
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v44
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v45
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v46
-; GCN-NEXT: v_add_f32_e32 v44, 0x40c00000, v47
-; GCN-NEXT: v_add_f32_e32 v45, 0x40c00000, v56
-; GCN-NEXT: v_add_f32_e32 v46, 0x40c00000, v57
-; GCN-NEXT: v_add_f32_e32 v47, 0x40c00000, v58
-; GCN-NEXT: v_add_f32_e32 v56, 0x40c00000, v59
-; GCN-NEXT: v_add_f32_e32 v57, 0x40c00000, v60
-; GCN-NEXT: v_add_f32_e32 v58, 0x40c00000, v61
-; GCN-NEXT: v_add_f32_e32 v59, 0x40c00000, v62
-; GCN-NEXT: v_add_f32_e32 v60, 0x40c00000, v63
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v61, 0x40c00000, v61
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v63, 0x40c00000, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 16, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 16, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 16, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 16, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v61
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v58
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v31
-; GCN-NEXT: v_mov_b32_e32 v31, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v29
-; GCN-NEXT: v_mov_b32_e32 v29, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v28, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v27, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v26, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v25, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v24, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v23, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v22, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v21, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v19, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v16, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v15, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v14, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v13, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v12, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v11, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v9, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v8, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v7, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v6, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v5, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v1
-; GCN-NEXT: .LBB50_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v4, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v2, v1
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v2, v1
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v2, v1
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 20, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v4, v4, v2
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 24, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v6, v6, v2
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 28, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v8, v8, v2
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v10, v10, v2
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 36, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v12, v12, v2
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 40, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v14, v14, v2
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v16, v16, v2
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 48, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v18, v18, v2
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v20, v20, v2
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v22, v2
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v25, v26, v25
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_or_b32_e32 v27, v28, v27
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_or_b32_e32 v29, v30, v29
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v31, v32, v31
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v33, v34, v33
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_or_b32_e32 v35, v36, v35
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v37, v38, v37
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_or_b32_e32 v39, v48, v39
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_or_b32_e32 v49, v50, v49
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_or_b32_e32 v51, v52, v51
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_or_b32_e32 v53, v54, v53
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_or_b32_e32 v55, v40, v55
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_or_b32_e32 v41, v42, v41
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v43, v3
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v45, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v56, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v30, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v51, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v53, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v55, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB50_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v17
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
-; VI-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; VI-NEXT: v_bfe_u32 v35, v34, 16, 1
-; VI-NEXT: v_add_u32_e32 v35, vcc, v35, v34
-; VI-NEXT: v_add_u32_e32 v35, vcc, s6, v35
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v34, v34
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc
-; VI-NEXT: v_bfe_u32 v35, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v35, vcc, v35, v18
-; VI-NEXT: v_add_u32_e32 v35, vcc, s6, v35
-; VI-NEXT: v_or_b32_e32 v36, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v35, v36, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
-; VI-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; VI-NEXT: v_bfe_u32 v36, v35, 16, 1
-; VI-NEXT: v_add_u32_e32 v36, vcc, v36, v35
-; VI-NEXT: v_add_u32_e32 v36, vcc, s6, v36
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v37, 0x400000, v35
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v35, v35
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v35, v36, v37, vcc
-; VI-NEXT: v_bfe_u32 v36, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v36, vcc, v36, v19
-; VI-NEXT: v_add_u32_e32 v36, vcc, s6, v36
-; VI-NEXT: v_or_b32_e32 v37, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v36, v37, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v20
-; VI-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; VI-NEXT: v_bfe_u32 v37, v36, 16, 1
-; VI-NEXT: v_add_u32_e32 v37, vcc, v37, v36
-; VI-NEXT: v_add_u32_e32 v37, vcc, s6, v37
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v36, v36
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v36, v37, v38, vcc
-; VI-NEXT: v_bfe_u32 v37, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v37, vcc, v37, v20
-; VI-NEXT: v_add_u32_e32 v37, vcc, s6, v37
-; VI-NEXT: v_or_b32_e32 v38, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v37, v38, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; VI-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
-; VI-NEXT: v_bfe_u32 v38, v37, 16, 1
-; VI-NEXT: v_add_u32_e32 v38, vcc, v38, v37
-; VI-NEXT: v_add_u32_e32 v38, vcc, s6, v38
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v37, v37
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v37, v38, v39, vcc
-; VI-NEXT: v_bfe_u32 v38, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v38, vcc, v38, v21
-; VI-NEXT: v_add_u32_e32 v38, vcc, s6, v38
-; VI-NEXT: v_or_b32_e32 v39, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v38, v39, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v22
-; VI-NEXT: v_add_f32_e32 v38, 0x40c00000, v38
-; VI-NEXT: v_bfe_u32 v39, v38, 16, 1
-; VI-NEXT: v_add_u32_e32 v39, vcc, v39, v38
-; VI-NEXT: v_add_u32_e32 v39, vcc, s6, v39
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v48, 0x400000, v38
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v38, v38
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v38, v39, v48, vcc
-; VI-NEXT: v_bfe_u32 v39, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v39, vcc, v39, v22
-; VI-NEXT: v_add_u32_e32 v39, vcc, s6, v39
-; VI-NEXT: v_or_b32_e32 v48, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v39, v48, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v23
-; VI-NEXT: v_add_f32_e32 v39, 0x40c00000, v39
-; VI-NEXT: v_bfe_u32 v48, v39, 16, 1
-; VI-NEXT: v_add_u32_e32 v48, vcc, v48, v39
-; VI-NEXT: v_add_u32_e32 v48, vcc, s6, v48
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v49, 0x400000, v39
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v39, v39
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v39, v48, v49, vcc
-; VI-NEXT: v_bfe_u32 v48, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v48, vcc, v48, v23
-; VI-NEXT: v_add_u32_e32 v48, vcc, s6, v48
-; VI-NEXT: v_or_b32_e32 v49, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v24
-; VI-NEXT: v_add_f32_e32 v48, 0x40c00000, v48
-; VI-NEXT: v_bfe_u32 v49, v48, 16, 1
-; VI-NEXT: v_add_u32_e32 v49, vcc, v49, v48
-; VI-NEXT: v_add_u32_e32 v49, vcc, s6, v49
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v50, 0x400000, v48
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v48, v48
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v48, v49, v50, vcc
-; VI-NEXT: v_bfe_u32 v49, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v49, vcc, v49, v24
-; VI-NEXT: v_add_u32_e32 v49, vcc, s6, v49
-; VI-NEXT: v_or_b32_e32 v50, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v49, v50, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v25
-; VI-NEXT: v_add_f32_e32 v49, 0x40c00000, v49
-; VI-NEXT: v_bfe_u32 v50, v49, 16, 1
-; VI-NEXT: v_add_u32_e32 v50, vcc, v50, v49
-; VI-NEXT: v_add_u32_e32 v50, vcc, s6, v50
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v51, 0x400000, v49
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v49, v49
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v49, v50, v51, vcc
-; VI-NEXT: v_bfe_u32 v50, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v50, vcc, v50, v25
-; VI-NEXT: v_add_u32_e32 v50, vcc, s6, v50
-; VI-NEXT: v_or_b32_e32 v51, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v50, v51, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v26
-; VI-NEXT: v_add_f32_e32 v50, 0x40c00000, v50
-; VI-NEXT: v_bfe_u32 v51, v50, 16, 1
-; VI-NEXT: v_add_u32_e32 v51, vcc, v51, v50
-; VI-NEXT: v_add_u32_e32 v51, vcc, s6, v51
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v52, 0x400000, v50
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v50, v50
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v50, v51, v52, vcc
-; VI-NEXT: v_bfe_u32 v51, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v51, vcc, v51, v26
-; VI-NEXT: v_add_u32_e32 v51, vcc, s6, v51
-; VI-NEXT: v_or_b32_e32 v52, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v51, v52, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
-; VI-NEXT: v_add_f32_e32 v51, 0x40c00000, v51
-; VI-NEXT: v_bfe_u32 v52, v51, 16, 1
-; VI-NEXT: v_add_u32_e32 v52, vcc, v52, v51
-; VI-NEXT: v_add_u32_e32 v52, vcc, s6, v52
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v53, 0x400000, v51
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v51, v51
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v51, v52, v53, vcc
-; VI-NEXT: v_bfe_u32 v52, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v52, vcc, v52, v27
-; VI-NEXT: v_add_u32_e32 v52, vcc, s6, v52
-; VI-NEXT: v_or_b32_e32 v53, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v52, v53, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v28
-; VI-NEXT: v_add_f32_e32 v52, 0x40c00000, v52
-; VI-NEXT: v_bfe_u32 v53, v52, 16, 1
-; VI-NEXT: v_add_u32_e32 v53, vcc, v53, v52
-; VI-NEXT: v_add_u32_e32 v53, vcc, s6, v53
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v54, 0x400000, v52
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v52, v52
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v52, v53, v54, vcc
-; VI-NEXT: v_bfe_u32 v53, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v53, vcc, v53, v28
-; VI-NEXT: v_add_u32_e32 v53, vcc, s6, v53
-; VI-NEXT: v_or_b32_e32 v54, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v29
-; VI-NEXT: v_add_f32_e32 v53, 0x40c00000, v53
-; VI-NEXT: v_bfe_u32 v54, v53, 16, 1
-; VI-NEXT: v_add_u32_e32 v54, vcc, v54, v53
-; VI-NEXT: v_add_u32_e32 v54, vcc, s6, v54
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v55, 0x400000, v53
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v53, v53
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v53, v54, v55, vcc
-; VI-NEXT: v_bfe_u32 v54, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v54, vcc, v54, v29
-; VI-NEXT: v_add_u32_e32 v54, vcc, s6, v54
-; VI-NEXT: v_or_b32_e32 v55, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v54, v55, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v30
-; VI-NEXT: v_add_f32_e32 v54, 0x40c00000, v54
-; VI-NEXT: v_bfe_u32 v55, v54, 16, 1
-; VI-NEXT: v_add_u32_e32 v55, vcc, v55, v54
-; VI-NEXT: v_add_u32_e32 v55, vcc, s6, v55
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v40, 0x400000, v54
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v54, v54
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v54, v55, v40, vcc
-; VI-NEXT: v_bfe_u32 v55, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v55, vcc, v55, v30
-; VI-NEXT: v_add_u32_e32 v55, vcc, s6, v55
-; VI-NEXT: v_or_b32_e32 v40, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v55, v40, vcc
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v31
-; VI-NEXT: v_add_f32_e32 v55, 0x40c00000, v55
-; VI-NEXT: v_bfe_u32 v40, v55, 16, 1
-; VI-NEXT: v_add_u32_e32 v40, vcc, v40, v55
-; VI-NEXT: v_add_u32_e32 v40, vcc, s6, v40
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v41, 0x400000, v55
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v55, v55
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v55, v40, v41, vcc
-; VI-NEXT: v_bfe_u32 v40, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v40, vcc, v40, v31
-; VI-NEXT: v_add_u32_e32 v40, vcc, s6, v40
-; VI-NEXT: v_or_b32_e32 v41, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v40, v41, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v0
-; VI-NEXT: v_add_f32_e32 v40, 0x40c00000, v40
-; VI-NEXT: v_bfe_u32 v41, v40, 16, 1
-; VI-NEXT: v_add_u32_e32 v41, vcc, v41, v40
-; VI-NEXT: v_add_u32_e32 v41, vcc, s6, v41
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v42, 0x400000, v40
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v40, v40
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v40, v41, v42, vcc
-; VI-NEXT: v_bfe_u32 v41, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v41, vcc, v41, v0
-; VI-NEXT: v_add_u32_e32 v41, vcc, s6, v41
-; VI-NEXT: v_or_b32_e32 v42, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v17, v41, v42, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v1
-; VI-NEXT: v_add_f32_e32 v41, 0x40c00000, v41
-; VI-NEXT: v_bfe_u32 v42, v41, 16, 1
-; VI-NEXT: v_add_u32_e32 v42, vcc, v42, v41
-; VI-NEXT: v_add_u32_e32 v42, vcc, s6, v42
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v43, 0x400000, v41
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v41, v41
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v41, v42, v43, vcc
-; VI-NEXT: v_bfe_u32 v42, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v42, vcc, v42, v1
-; VI-NEXT: v_add_u32_e32 v42, vcc, s6, v42
-; VI-NEXT: v_or_b32_e32 v43, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_cndmask_b32_e32 v16, v42, v43, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v2
-; VI-NEXT: v_add_f32_e32 v42, 0x40c00000, v42
-; VI-NEXT: v_bfe_u32 v43, v42, 16, 1
-; VI-NEXT: v_add_u32_e32 v43, vcc, v43, v42
-; VI-NEXT: v_add_u32_e32 v43, vcc, s6, v43
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v44, 0x400000, v42
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v42, v42
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v42, v43, v44, vcc
-; VI-NEXT: v_bfe_u32 v43, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v43, vcc, v43, v2
-; VI-NEXT: v_add_u32_e32 v43, vcc, s6, v43
-; VI-NEXT: v_or_b32_e32 v44, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v43, v44, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v3
-; VI-NEXT: v_add_f32_e32 v43, 0x40c00000, v43
-; VI-NEXT: v_bfe_u32 v44, v43, 16, 1
-; VI-NEXT: v_add_u32_e32 v44, vcc, v44, v43
-; VI-NEXT: v_add_u32_e32 v44, vcc, s6, v44
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v45, 0x400000, v43
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v43, v43
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v43, v44, v45, vcc
-; VI-NEXT: v_bfe_u32 v44, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v44, vcc, v44, v3
-; VI-NEXT: v_add_u32_e32 v44, vcc, s6, v44
-; VI-NEXT: v_or_b32_e32 v45, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v44, v45, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v4
-; VI-NEXT: v_add_f32_e32 v44, 0x40c00000, v44
-; VI-NEXT: v_bfe_u32 v45, v44, 16, 1
-; VI-NEXT: v_add_u32_e32 v45, vcc, v45, v44
-; VI-NEXT: v_add_u32_e32 v45, vcc, s6, v45
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v46, 0x400000, v44
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v44, v44
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v44, v45, v46, vcc
-; VI-NEXT: v_bfe_u32 v45, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v45, vcc, v45, v4
-; VI-NEXT: v_add_u32_e32 v45, vcc, s6, v45
-; VI-NEXT: v_or_b32_e32 v46, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v45, v46, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v5
-; VI-NEXT: v_add_f32_e32 v45, 0x40c00000, v45
-; VI-NEXT: v_bfe_u32 v46, v45, 16, 1
-; VI-NEXT: v_add_u32_e32 v46, vcc, v46, v45
-; VI-NEXT: v_add_u32_e32 v46, vcc, s6, v46
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v47, 0x400000, v45
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v45, v45
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v45, v46, v47, vcc
-; VI-NEXT: v_bfe_u32 v46, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v46, vcc, v46, v5
-; VI-NEXT: v_add_u32_e32 v46, vcc, s6, v46
-; VI-NEXT: v_or_b32_e32 v47, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v46, v47, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; VI-NEXT: v_add_f32_e32 v46, 0x40c00000, v46
-; VI-NEXT: v_bfe_u32 v47, v46, 16, 1
-; VI-NEXT: v_add_u32_e32 v47, vcc, v47, v46
-; VI-NEXT: v_add_u32_e32 v47, vcc, s6, v47
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v56, 0x400000, v46
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v46, v46
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v46, v47, v56, vcc
-; VI-NEXT: v_bfe_u32 v47, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v47, vcc, v47, v6
-; VI-NEXT: v_add_u32_e32 v47, vcc, s6, v47
-; VI-NEXT: v_or_b32_e32 v56, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v47, v56, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v7
-; VI-NEXT: v_add_f32_e32 v47, 0x40c00000, v47
-; VI-NEXT: v_bfe_u32 v56, v47, 16, 1
-; VI-NEXT: v_add_u32_e32 v56, vcc, v56, v47
-; VI-NEXT: v_add_u32_e32 v56, vcc, s6, v56
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v57, 0x400000, v47
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v47, v47
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v47, v56, v57, vcc
-; VI-NEXT: v_bfe_u32 v56, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v56, vcc, v56, v7
-; VI-NEXT: v_add_u32_e32 v56, vcc, s6, v56
-; VI-NEXT: v_or_b32_e32 v57, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v56, v57, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v56, 16, v8
-; VI-NEXT: v_add_f32_e32 v56, 0x40c00000, v56
-; VI-NEXT: v_bfe_u32 v57, v56, 16, 1
-; VI-NEXT: v_add_u32_e32 v57, vcc, v57, v56
-; VI-NEXT: v_add_u32_e32 v57, vcc, s6, v57
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v58, 0x400000, v56
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v56, v56
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v56, v57, v58, vcc
-; VI-NEXT: v_bfe_u32 v57, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v57, vcc, v57, v8
-; VI-NEXT: v_add_u32_e32 v57, vcc, s6, v57
-; VI-NEXT: v_or_b32_e32 v58, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v57, v58, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v57, 16, v9
-; VI-NEXT: v_add_f32_e32 v57, 0x40c00000, v57
-; VI-NEXT: v_bfe_u32 v58, v57, 16, 1
-; VI-NEXT: v_add_u32_e32 v58, vcc, v58, v57
-; VI-NEXT: v_add_u32_e32 v58, vcc, s6, v58
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v59, 0x400000, v57
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v57, v57
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v57, v58, v59, vcc
-; VI-NEXT: v_bfe_u32 v58, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v58, vcc, v58, v9
-; VI-NEXT: v_add_u32_e32 v58, vcc, s6, v58
-; VI-NEXT: v_or_b32_e32 v59, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v58, v59, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v58, 16, v10
-; VI-NEXT: v_add_f32_e32 v58, 0x40c00000, v58
-; VI-NEXT: v_bfe_u32 v59, v58, 16, 1
-; VI-NEXT: v_add_u32_e32 v59, vcc, v59, v58
-; VI-NEXT: v_add_u32_e32 v59, vcc, s6, v59
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v60, 0x400000, v58
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v58, v58
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v58, v59, v60, vcc
-; VI-NEXT: v_bfe_u32 v59, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v59, vcc, v59, v10
-; VI-NEXT: v_add_u32_e32 v59, vcc, s6, v59
-; VI-NEXT: v_or_b32_e32 v60, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v59, v60, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v59, 16, v11
-; VI-NEXT: v_add_f32_e32 v59, 0x40c00000, v59
-; VI-NEXT: v_bfe_u32 v60, v59, 16, 1
-; VI-NEXT: v_add_u32_e32 v60, vcc, v60, v59
-; VI-NEXT: v_add_u32_e32 v60, vcc, s6, v60
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v61, 0x400000, v59
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v59, v59
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v59, v60, v61, vcc
-; VI-NEXT: v_bfe_u32 v60, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v60, vcc, v60, v11
-; VI-NEXT: v_add_u32_e32 v60, vcc, s6, v60
-; VI-NEXT: v_or_b32_e32 v61, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v60, v61, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v60, 16, v12
-; VI-NEXT: v_add_f32_e32 v60, 0x40c00000, v60
-; VI-NEXT: v_bfe_u32 v61, v60, 16, 1
-; VI-NEXT: v_add_u32_e32 v61, vcc, v61, v60
-; VI-NEXT: v_add_u32_e32 v61, vcc, s6, v61
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v62, 0x400000, v60
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v60, v60
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v60, v61, v62, vcc
-; VI-NEXT: v_bfe_u32 v61, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v61, vcc, v61, v12
-; VI-NEXT: v_add_u32_e32 v61, vcc, s6, v61
-; VI-NEXT: v_or_b32_e32 v62, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v61, v62, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v61, 16, v13
-; VI-NEXT: v_add_f32_e32 v61, 0x40c00000, v61
-; VI-NEXT: v_bfe_u32 v62, v61, 16, 1
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v61
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v61
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v61, v61
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v61, v62, v63, vcc
-; VI-NEXT: v_bfe_u32 v62, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v13
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v62, v63, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v62, 16, v14
-; VI-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; VI-NEXT: v_bfe_u32 v63, v62, 16, 1
-; VI-NEXT: v_add_u32_e32 v63, vcc, v63, v62
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_add_u32_e32 v63, vcc, s6, v63
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; VI-NEXT: v_bfe_u32 v62, v14, 16, 1
-; VI-NEXT: v_cndmask_b32_e32 v1, v63, v0, vcc
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v14
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v62, v63, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v62, 16, v15
-; VI-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; VI-NEXT: v_bfe_u32 v63, v62, 16, 1
-; VI-NEXT: v_add_u32_e32 v63, vcc, v63, v62
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_add_u32_e32 v63, vcc, s6, v63
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; VI-NEXT: v_bfe_u32 v62, v15, 16, 1
-; VI-NEXT: v_cndmask_b32_e32 v0, v63, v0, vcc
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v15
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v62, v63, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v0, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v0, v1, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v0, v61, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v0, v60, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v0, v59, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v0, v58, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v0, v57, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v0, v56, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v0, v47, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v0, v46, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v0, v45, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v0, v44, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v0, v43, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v0, v42, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v16, v55, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v16, v54, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v16, v53, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v16, v52, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v16, v51, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v16, v50, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v16, v49, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v16, v48, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v16, v39, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v16, v38, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v16, v37, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v16, v36, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v16, v35, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v16, v34, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v32
-; VI-NEXT: v_alignbit_b32 v1, v0, v41, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v16, v33, 16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: v_alignbit_b32 v0, v0, v40, 16
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB50_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB50_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; GFX9-NEXT: v_bfe_u32 v34, v33, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v34, v34, v33, s6
-; GFX9-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; GFX9-NEXT: v_bfe_u32 v34, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v34, v34, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v35, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX9-NEXT: v_bfe_u32 v35, v34, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v35, v35, v34, s6
-; GFX9-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v34, v34
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc
-; GFX9-NEXT: v_bfe_u32 v35, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v35, v35, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v36, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v35, v36, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v35, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX9-NEXT: v_bfe_u32 v36, v35, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v36, v36, v35, s6
-; GFX9-NEXT: v_or_b32_e32 v37, 0x400000, v35
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v35, v35
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v35, v36, v37, vcc
-; GFX9-NEXT: v_bfe_u32 v36, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v36, v36, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v37, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v36, v37, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v36, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX9-NEXT: v_bfe_u32 v37, v36, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v37, v37, v36, s6
-; GFX9-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v36, v36
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v36, v37, v38, vcc
-; GFX9-NEXT: v_bfe_u32 v37, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v37, v37, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v38, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v37, v38, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
-; GFX9-NEXT: v_bfe_u32 v38, v37, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v38, v38, v37, s6
-; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v37, v37
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v37, v38, v39, vcc
-; GFX9-NEXT: v_bfe_u32 v38, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v38, v38, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v38, v39, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v38, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v38, 0x40c00000, v38
-; GFX9-NEXT: v_bfe_u32 v39, v38, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v39, v39, v38, s6
-; GFX9-NEXT: v_or_b32_e32 v48, 0x400000, v38
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v38, v38
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v38, v39, v48, vcc
-; GFX9-NEXT: v_bfe_u32 v39, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v39, v39, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v48, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v39, v48, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v39, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v39, 0x40c00000, v39
-; GFX9-NEXT: v_bfe_u32 v48, v39, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v48, v48, v39, s6
-; GFX9-NEXT: v_or_b32_e32 v49, 0x400000, v39
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v39, v39
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v39, v48, v49, vcc
-; GFX9-NEXT: v_bfe_u32 v48, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v48, v48, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v49, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v48, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v48, 0x40c00000, v48
-; GFX9-NEXT: v_bfe_u32 v49, v48, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v49, v49, v48, s6
-; GFX9-NEXT: v_or_b32_e32 v50, 0x400000, v48
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v48, v48
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v48, v49, v50, vcc
-; GFX9-NEXT: v_bfe_u32 v49, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v49, v49, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v50, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v49, v50, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v49, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v49, 0x40c00000, v49
-; GFX9-NEXT: v_bfe_u32 v50, v49, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v50, v50, v49, s6
-; GFX9-NEXT: v_or_b32_e32 v51, 0x400000, v49
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v49, v49
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v49, v50, v51, vcc
-; GFX9-NEXT: v_bfe_u32 v50, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v50, v50, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v51, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v50, v51, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v50, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v50, 0x40c00000, v50
-; GFX9-NEXT: v_bfe_u32 v51, v50, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v51, v51, v50, s6
-; GFX9-NEXT: v_or_b32_e32 v52, 0x400000, v50
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v50, v50
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v50, v51, v52, vcc
-; GFX9-NEXT: v_bfe_u32 v51, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v51, v51, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v52, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v51, v52, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v51, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v51, 0x40c00000, v51
-; GFX9-NEXT: v_bfe_u32 v52, v51, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v52, v52, v51, s6
-; GFX9-NEXT: v_or_b32_e32 v53, 0x400000, v51
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v51, v51
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v51, v52, v53, vcc
-; GFX9-NEXT: v_bfe_u32 v52, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v52, v52, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v53, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v52, v53, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v52, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v52, 0x40c00000, v52
-; GFX9-NEXT: v_bfe_u32 v53, v52, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v53, v53, v52, s6
-; GFX9-NEXT: v_or_b32_e32 v54, 0x400000, v52
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v52, v52
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v52, v53, v54, vcc
-; GFX9-NEXT: v_bfe_u32 v53, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v53, v53, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v54, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v53, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v53, 0x40c00000, v53
-; GFX9-NEXT: v_bfe_u32 v54, v53, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v54, v54, v53, s6
-; GFX9-NEXT: v_or_b32_e32 v55, 0x400000, v53
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v53, v53
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v53, v54, v55, vcc
-; GFX9-NEXT: v_bfe_u32 v54, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v54, v54, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v55, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v54, v55, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v54, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v54, 0x40c00000, v54
-; GFX9-NEXT: v_bfe_u32 v55, v54, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v55, v55, v54, s6
-; GFX9-NEXT: v_or_b32_e32 v40, 0x400000, v54
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v54, v54
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v54, v55, v40, vcc
-; GFX9-NEXT: v_bfe_u32 v55, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v55, v55, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v40, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v55, v40, vcc
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v55, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v55, 0x40c00000, v55
-; GFX9-NEXT: v_bfe_u32 v40, v55, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v40, v40, v55, s6
-; GFX9-NEXT: v_or_b32_e32 v41, 0x400000, v55
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v55, v55
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v55, v40, v41, vcc
-; GFX9-NEXT: v_bfe_u32 v40, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v40, v40, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v41, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v40, v41, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v40, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v40, 0x40c00000, v40
-; GFX9-NEXT: v_bfe_u32 v41, v40, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v41, v41, v40, s6
-; GFX9-NEXT: v_or_b32_e32 v42, 0x400000, v40
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v40, v40
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v40, v41, v42, vcc
-; GFX9-NEXT: v_bfe_u32 v41, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v41, v41, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v42, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v41, v42, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v41, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v41, 0x40c00000, v41
-; GFX9-NEXT: v_bfe_u32 v42, v41, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v42, v42, v41, s6
-; GFX9-NEXT: v_or_b32_e32 v43, 0x400000, v41
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v41, v41
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v41, v42, v43, vcc
-; GFX9-NEXT: v_bfe_u32 v42, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v42, v42, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v43, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v42, v43, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v42, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v42, 0x40c00000, v42
-; GFX9-NEXT: v_bfe_u32 v43, v42, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v43, v43, v42, s6
-; GFX9-NEXT: v_or_b32_e32 v44, 0x400000, v42
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v42, v42
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v42, v43, v44, vcc
-; GFX9-NEXT: v_bfe_u32 v43, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v43, v43, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v44, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v43, v44, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v43, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v43, 0x40c00000, v43
-; GFX9-NEXT: v_bfe_u32 v44, v43, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v44, v44, v43, s6
-; GFX9-NEXT: v_or_b32_e32 v45, 0x400000, v43
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v43, v43
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v43, v44, v45, vcc
-; GFX9-NEXT: v_bfe_u32 v44, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v44, v44, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v45, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v44, v45, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v44, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v44, 0x40c00000, v44
-; GFX9-NEXT: v_bfe_u32 v45, v44, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v45, v45, v44, s6
-; GFX9-NEXT: v_or_b32_e32 v46, 0x400000, v44
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v44, v44
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v44, v45, v46, vcc
-; GFX9-NEXT: v_bfe_u32 v45, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v45, v45, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v46, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v45, v46, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v45, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v45, 0x40c00000, v45
-; GFX9-NEXT: v_bfe_u32 v46, v45, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v46, v46, v45, s6
-; GFX9-NEXT: v_or_b32_e32 v47, 0x400000, v45
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v45, v45
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v45, v46, v47, vcc
-; GFX9-NEXT: v_bfe_u32 v46, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v46, v46, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v47, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v46, v47, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v46, 0x40c00000, v46
-; GFX9-NEXT: v_bfe_u32 v47, v46, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v47, v47, v46, s6
-; GFX9-NEXT: v_or_b32_e32 v56, 0x400000, v46
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v46, v46
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v46, v47, v56, vcc
-; GFX9-NEXT: v_bfe_u32 v47, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v47, v47, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v56, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v47, v56, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v47, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v47, 0x40c00000, v47
-; GFX9-NEXT: v_bfe_u32 v56, v47, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v56, v56, v47, s6
-; GFX9-NEXT: v_or_b32_e32 v57, 0x400000, v47
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v47, v47
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v47, v56, v57, vcc
-; GFX9-NEXT: v_bfe_u32 v56, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v56, v56, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v57, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v56, v57, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v56, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v56, 0x40c00000, v56
-; GFX9-NEXT: v_bfe_u32 v57, v56, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v57, v57, v56, s6
-; GFX9-NEXT: v_or_b32_e32 v58, 0x400000, v56
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v56, v56
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v56, v57, v58, vcc
-; GFX9-NEXT: v_bfe_u32 v57, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v57, v57, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v58, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v57, v58, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v57, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v57, 0x40c00000, v57
-; GFX9-NEXT: v_bfe_u32 v58, v57, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v58, v58, v57, s6
-; GFX9-NEXT: v_or_b32_e32 v59, 0x400000, v57
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v57, v57
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v57, v58, v59, vcc
-; GFX9-NEXT: v_bfe_u32 v58, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v58, v58, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v59, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v58, v59, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v58, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v58, 0x40c00000, v58
-; GFX9-NEXT: v_bfe_u32 v59, v58, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v59, v59, v58, s6
-; GFX9-NEXT: v_or_b32_e32 v60, 0x400000, v58
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v58, v58
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v58, v59, v60, vcc
-; GFX9-NEXT: v_bfe_u32 v59, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v59, v59, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v60, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v59, v60, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v59, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v59, 0x40c00000, v59
-; GFX9-NEXT: v_bfe_u32 v60, v59, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v60, v60, v59, s6
-; GFX9-NEXT: v_or_b32_e32 v61, 0x400000, v59
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v59, v59
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v59, v60, v61, vcc
-; GFX9-NEXT: v_bfe_u32 v60, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v60, v60, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v61, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v60, v61, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v60, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v60, 0x40c00000, v60
-; GFX9-NEXT: v_bfe_u32 v61, v60, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v61, v61, v60, s6
-; GFX9-NEXT: v_or_b32_e32 v62, 0x400000, v60
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v60, v60
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v60, v61, v62, vcc
-; GFX9-NEXT: v_bfe_u32 v61, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v61, v61, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v62, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v61, v62, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v61, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v61, 0x40c00000, v61
-; GFX9-NEXT: v_bfe_u32 v62, v61, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v62, v62, v61, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v61
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v61, v61
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v61, v62, v63, vcc
-; GFX9-NEXT: v_bfe_u32 v62, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v62, v62, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v62, v63, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v62, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_bfe_u32 v63, v62, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_add3_u32 v63, v63, v62, s6
-; GFX9-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; GFX9-NEXT: v_bfe_u32 v62, v14, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v63, v0, vcc
-; GFX9-NEXT: v_add3_u32 v62, v62, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v62, v63, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v62, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_bfe_u32 v63, v62, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_add3_u32 v63, v63, v62, s6
-; GFX9-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; GFX9-NEXT: v_bfe_u32 v62, v15, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v63, v0, vcc
-; GFX9-NEXT: v_add3_u32 v62, v62, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v62, v63, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v0, s6
-; GFX9-NEXT: v_perm_b32 v14, v14, v1, s6
-; GFX9-NEXT: v_perm_b32 v1, v16, v41, s6
-; GFX9-NEXT: v_perm_b32 v0, v17, v40, s6
-; GFX9-NEXT: v_perm_b32 v17, v32, v33, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v13, v13, v61, s6
-; GFX9-NEXT: v_perm_b32 v12, v12, v60, s6
-; GFX9-NEXT: v_perm_b32 v11, v11, v59, s6
-; GFX9-NEXT: v_perm_b32 v10, v10, v58, s6
-; GFX9-NEXT: v_perm_b32 v9, v9, v57, s6
-; GFX9-NEXT: v_perm_b32 v8, v8, v56, s6
-; GFX9-NEXT: v_perm_b32 v7, v7, v47, s6
-; GFX9-NEXT: v_perm_b32 v6, v6, v46, s6
-; GFX9-NEXT: v_perm_b32 v5, v5, v45, s6
-; GFX9-NEXT: v_perm_b32 v4, v4, v44, s6
-; GFX9-NEXT: v_perm_b32 v3, v3, v43, s6
-; GFX9-NEXT: v_perm_b32 v2, v2, v42, s6
-; GFX9-NEXT: v_perm_b32 v31, v31, v55, s6
-; GFX9-NEXT: v_perm_b32 v30, v30, v54, s6
-; GFX9-NEXT: v_perm_b32 v29, v29, v53, s6
-; GFX9-NEXT: v_perm_b32 v28, v28, v52, s6
-; GFX9-NEXT: v_perm_b32 v27, v27, v51, s6
-; GFX9-NEXT: v_perm_b32 v26, v26, v50, s6
-; GFX9-NEXT: v_perm_b32 v25, v25, v49, s6
-; GFX9-NEXT: v_perm_b32 v24, v24, v48, s6
-; GFX9-NEXT: v_perm_b32 v23, v23, v39, s6
-; GFX9-NEXT: v_perm_b32 v22, v22, v38, s6
-; GFX9-NEXT: v_perm_b32 v21, v21, v37, s6
-; GFX9-NEXT: v_perm_b32 v20, v20, v36, s6
-; GFX9-NEXT: v_perm_b32 v19, v19, v35, s6
-; GFX9-NEXT: v_perm_b32 v18, v18, v34, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v16, v32, v16, s6
-; GFX9-NEXT: .LBB50_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB50_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_dual_add_f32 v69, 0x40c00000, v69 :: v_dual_lshlrev_b32 v70, 16, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v32
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v10
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v87, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v16
-; GFX11-NEXT: v_bfe_u32 v16, v32, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v87, 0x40c00000, v87
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v16, v16, v32, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v9
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v38, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v17
-; GFX11-NEXT: v_bfe_u32 v17, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v38, 0x40c00000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v11
-; GFX11-NEXT: v_add3_u32 v17, v17, v34, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v34, v36, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v17, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v33
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_add3_u32 v33, v34, v36, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v37, v17, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v37, v35, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v39, 0x40c00000, v18
-; GFX11-NEXT: v_add3_u32 v34, v37, v35, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v37, v38, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_perm_b32 v17, v33, v17, 0x7060302
-; GFX11-NEXT: v_perm_b32 v16, v32, v16, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v34, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v34, v37, v38, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v36, v39, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v37, 0x40c00000, v19
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v37, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v36, v39, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v39
-; GFX11-NEXT: v_add_f32_e32 v48, 0x40c00000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_add_f32_e32 v39, 0x40c00000, v20
-; GFX11-NEXT: v_perm_b32 v18, v34, v18, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v38, v37, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v48, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v48
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v36, v38, v48, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v39, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_dual_add_f32 v49, 0x40c00000, v20 :: v_dual_add_f32 v48, 0x40c00000, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v36, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v36, v38, v39, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v39
-; GFX11-NEXT: v_bfe_u32 v38, v49, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_bfe_u32 v39, v48, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v19, v35, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v36, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v38, v49, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v49
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_dual_add_f32 v49, 0x40c00000, v22 :: v_dual_lshlrev_b32 v22, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_perm_b32 v20, v36, v20, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v39, v48, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v48
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_bfe_u32 v48, v49, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v51, 0x40c00000, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v37, v37, v38, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v39, v50, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_perm_b32 v21, v37, v21, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v38, v39, v50, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v50
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v22, v38, v39, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v48, v49, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v49
-; GFX11-NEXT: v_bfe_u32 v48, v51, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v24
-; GFX11-NEXT: v_bfe_u32 v49, v50, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_cndmask_b32_e32 v38, v38, v39, vcc_lo
-; GFX11-NEXT: v_add3_u32 v39, v48, v51, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v51
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: v_dual_add_f32 v52, 0x40c00000, v23 :: v_dual_add_f32 v51, 0x40c00000, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v39, v48, vcc_lo
-; GFX11-NEXT: v_add3_u32 v39, v49, v50, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v50
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_bfe_u32 v50, v51, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v53, 0x40c00000, v24
-; GFX11-NEXT: v_perm_b32 v22, v38, v22, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v39, v39, v48, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v49, v52, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v23, v39, v23, 0x7060302
-; GFX11-NEXT: v_add3_u32 v48, v49, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v52
-; GFX11-NEXT: v_add_f32_e32 v52, 0x40c00000, v25
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v48, v49, vcc_lo
-; GFX11-NEXT: v_add3_u32 v48, v50, v51, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v51
-; GFX11-NEXT: v_bfe_u32 v50, v53, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v26
-; GFX11-NEXT: v_bfe_u32 v51, v52, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_cndmask_b32_e32 v48, v48, v49, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v50, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v53
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: v_dual_add_f32 v54, 0x40c00000, v25 :: v_dual_add_f32 v53, 0x40c00000, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v49, v50, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v51, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v52
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_bfe_u32 v52, v53, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v55, 0x40c00000, v26
-; GFX11-NEXT: v_perm_b32 v24, v48, v24, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v49, v49, v50, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v51, v54, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v54, v54
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v25, v49, v25, 0x7060302
-; GFX11-NEXT: v_add3_u32 v50, v51, v54, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v54
-; GFX11-NEXT: v_add_f32_e32 v54, 0x40c00000, v27
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v26, v50, v51, vcc_lo
-; GFX11-NEXT: v_add3_u32 v50, v52, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v53
-; GFX11-NEXT: v_bfe_u32 v52, v55, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GFX11-NEXT: v_bfe_u32 v53, v54, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_cndmask_b32_e32 v50, v50, v51, vcc_lo
-; GFX11-NEXT: v_add3_u32 v51, v52, v55, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v55
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55
-; GFX11-NEXT: v_dual_add_f32 v64, 0x40c00000, v27 :: v_dual_add_f32 v55, 0x40c00000, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v51, v52, vcc_lo
-; GFX11-NEXT: v_add3_u32 v51, v53, v54, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v54
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v54, v54
-; GFX11-NEXT: v_bfe_u32 v54, v55, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v65, 0x40c00000, v28
-; GFX11-NEXT: v_perm_b32 v26, v50, v26, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v51, v51, v52, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v53, v64, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v64, v64
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v27, v51, v27, 0x7060302
-; GFX11-NEXT: v_add3_u32 v52, v53, v64, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v64
-; GFX11-NEXT: v_add_f32_e32 v64, 0x40c00000, v29
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v52, v53, vcc_lo
-; GFX11-NEXT: v_add3_u32 v52, v54, v55, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v55
-; GFX11-NEXT: v_bfe_u32 v54, v65, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v30
-; GFX11-NEXT: v_bfe_u32 v55, v64, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_cndmask_b32_e32 v52, v52, v53, vcc_lo
-; GFX11-NEXT: v_add3_u32 v53, v54, v65, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v65
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v65, v65
-; GFX11-NEXT: v_dual_add_f32 v66, 0x40c00000, v29 :: v_dual_add_f32 v65, 0x40c00000, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v31
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v53, v54, vcc_lo
-; GFX11-NEXT: v_add3_u32 v53, v55, v64, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v64
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v64, v64
-; GFX11-NEXT: v_bfe_u32 v64, v65, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v67, 0x40c00000, v30
-; GFX11-NEXT: v_perm_b32 v28, v52, v28, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v53, v53, v54, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v55, v66, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v66, v66
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v29, v53, v29, 0x7060302
-; GFX11-NEXT: v_add3_u32 v54, v55, v66, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v66
-; GFX11-NEXT: v_add_f32_e32 v66, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v30, v54, v55, vcc_lo
-; GFX11-NEXT: v_add3_u32 v54, v64, v65, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v65
-; GFX11-NEXT: v_bfe_u32 v64, v67, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v65, v65
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v0
-; GFX11-NEXT: v_bfe_u32 v65, v66, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v54, v54, v55, vcc_lo
-; GFX11-NEXT: v_add3_u32 v55, v64, v67, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v67
-; GFX11-NEXT: v_add_f32_e32 v68, 0x40c00000, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v67, v67
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v67, 16, v1
-; GFX11-NEXT: v_perm_b32 v30, v54, v30, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v55, v64, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v67, 0x40c00000, v67
-; GFX11-NEXT: v_add3_u32 v55, v65, v66, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v66
-; GFX11-NEXT: v_bfe_u32 v65, v68, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v66, v66
-; GFX11-NEXT: v_bfe_u32 v66, v0, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v55, v55, v64, vcc_lo
-; GFX11-NEXT: v_add3_u32 v64, v65, v68, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v65, 0x400000, v68
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v68, v68
-; GFX11-NEXT: v_bfe_u32 v68, v67, 16, 1
-; GFX11-NEXT: v_perm_b32 v31, v55, v31, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v64, v64, v65, vcc_lo
-; GFX11-NEXT: v_add3_u32 v65, v66, v0, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_dual_cndmask_b32 v0, v65, v66 :: v_dual_and_b32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_add3_u32 v65, v68, v67, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v67
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v67, v67
-; GFX11-NEXT: v_perm_b32 v0, v0, v64, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v68, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v1
-; GFX11-NEXT: v_cndmask_b32_e32 v65, v65, v66, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v66, v68, v1, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v68, v69, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v66, v67, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v66, v68, v69, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v69
-; GFX11-NEXT: v_bfe_u32 v68, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v69, v69
-; GFX11-NEXT: v_perm_b32 v1, v1, v65, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v66, v66, v67, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v67, v68, v2, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v67, v68 :: v_dual_and_b32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v70, 0x40c00000, v70 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v2, v2, v66, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v69, v70, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v70
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v70, v70
-; GFX11-NEXT: v_add3_u32 v67, v69, v70, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v69, v3, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v67, v67, v68, vcc_lo
-; GFX11-NEXT: v_add3_u32 v68, v69, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v69, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v68, v69 :: v_dual_and_b32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_dual_add_f32 v71, 0x40c00000, v71 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v3, v3, v67, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v70, v71, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v69, 0x400000, v71
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v71, v71
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v68, v70, v71, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v70, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v68, v68, v69, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v69, v70, v4, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v69, v70 :: v_dual_and_b32 v5, 0xffff0000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_perm_b32 v4, v4, v68, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v71, v80, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v80
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v80, v80
-; GFX11-NEXT: v_add3_u32 v69, v71, v80, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v71, v5, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v69, v69, v70, vcc_lo
-; GFX11-NEXT: v_add3_u32 v70, v71, v5, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v71, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v70, v71 :: v_dual_and_b32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_dual_add_f32 v81, 0x40c00000, v81 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v5, v5, v69, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v80, v81, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v71, 0x400000, v81
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v81, v81
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v70, v80, v81, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v80, v6, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v70, v70, v71, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v71, v80, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v71, v80 :: v_dual_and_b32 v7, 0xffff0000, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_perm_b32 v6, v6, v70, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v81, v82, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v82
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-NEXT: v_add3_u32 v71, v81, v82, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v81, v7, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v71, v71, v80, vcc_lo
-; GFX11-NEXT: v_add3_u32 v80, v81, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v81, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v80, v81 :: v_dual_and_b32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_dual_add_f32 v83, 0x40c00000, v83 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v7, v7, v71, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v82, v83, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v81, 0x400000, v83
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v80, v82, v83, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v82, v8, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v80, v80, v81, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v81, v82, v8, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_dual_cndmask_b32 v8, v81, v82 :: v_dual_and_b32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v84, 0x40c00000, v84 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_perm_b32 v8, v8, v80, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v83, v84, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v84
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v84, v84
-; GFX11-NEXT: v_add3_u32 v81, v83, v84, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v83, v9, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v81, v81, v82, vcc_lo
-; GFX11-NEXT: v_add3_u32 v82, v83, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v83, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v82, v83 :: v_dual_and_b32 v10, 0xffff0000, v10
-; GFX11-NEXT: v_dual_add_f32 v85, 0x40c00000, v85 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v9, v9, v81, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v84, v85, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v83, 0x400000, v85
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v85, v85
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v82, v84, v85, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v84, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v82, v82, v83, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v83, v84, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v83, v84 :: v_dual_and_b32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v86, 0x40c00000, v86 :: v_dual_add_f32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_perm_b32 v10, v10, v82, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v85, v86, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v86
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v86, v86
-; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v11
-; GFX11-NEXT: v_add3_u32 v83, v85, v86, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v85, v11, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v83, v83, v84 :: v_dual_lshlrev_b32 v84, 16, v13
-; GFX11-NEXT: v_bfe_u32 v86, v87, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_or_b32_e32 v97, 0x400000, v87
-; GFX11-NEXT: v_add_f32_e32 v84, 0x40c00000, v84
-; GFX11-NEXT: v_add3_u32 v86, v86, v87, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v87, v87
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_bfe_u32 v98, v12, 16, 1
-; GFX11-NEXT: v_bfe_u32 v99, v84, 16, 1
-; GFX11-NEXT: v_add3_u32 v85, v85, v11, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v86, v86, v97, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v97, 0x400000, v12
-; GFX11-NEXT: v_add3_u32 v87, v98, v12, 0x7fff
-; GFX11-NEXT: v_add3_u32 v98, v99, v84, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v99, 16, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v84
-; GFX11-NEXT: v_bfe_u32 v101, v13, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v12, v87, v97 :: v_dual_add_f32 v87, 0x40c00000, v99
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v84, v84
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v97, v101, v13, 0x7fff
-; GFX11-NEXT: v_perm_b32 v12, v12, v86, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v99, v87, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v84, v98, v100, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v98, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v101, 0x400000, v87
-; GFX11-NEXT: v_bfe_u32 v102, v14, 16, 1
-; GFX11-NEXT: v_add3_u32 v99, v99, v87, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v87, v87
-; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v98, 0x40c00000, v98 :: v_dual_cndmask_b32 v87, v99, v101
-; GFX11-NEXT: v_add3_u32 v101, v102, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_bfe_u32 v103, v98, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v98
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v101, v102 :: v_dual_add_f32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v103, v103, v98, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v98, v98
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v14, v14, v87, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v99, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v113, 0x400000, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v98, v103, v112, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v99, v99, v15, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v99, v113, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v15, v15, v98, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v97, v100, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_perm_b32 v13, v13, v84, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-NEXT: .LBB50_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
-
-define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:32
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:120
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v45
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v56
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:112
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v57
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v31
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v1
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v60
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v61
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v62
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v31
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v2
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v7
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v11
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v12
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v13
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v14
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v15
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v18
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v19
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v21
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v23
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v27
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v34
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v36
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v37
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v38
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v39
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v48
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v49
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v50
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v51
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v52
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v53
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v54
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v55
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v40
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v42
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v43
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v44
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v45
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v46
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v47
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v56
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v57
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v58
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v60
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v61
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v62
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: .LBB51_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f32_f16_e32 v63, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v62, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v63
-; GCN-NEXT: v_add_f32_e32 v62, 0x38000000, v62
-; GCN-NEXT: v_add_f32_e32 v61, 0x38000000, v61
-; GCN-NEXT: v_add_f32_e32 v60, 0x38000000, v60
-; GCN-NEXT: v_add_f32_e32 v59, 0x38000000, v59
-; GCN-NEXT: v_add_f32_e32 v58, 0x38000000, v58
-; GCN-NEXT: v_add_f32_e32 v57, 0x38000000, v57
-; GCN-NEXT: v_add_f32_e32 v56, 0x38000000, v56
-; GCN-NEXT: v_add_f32_e32 v47, 0x38000000, v47
-; GCN-NEXT: v_add_f32_e32 v46, 0x38000000, v46
-; GCN-NEXT: v_add_f32_e32 v45, 0x38000000, v45
-; GCN-NEXT: v_add_f32_e32 v44, 0x38000000, v44
-; GCN-NEXT: v_add_f32_e32 v43, 0x38000000, v43
-; GCN-NEXT: v_add_f32_e32 v42, 0x38000000, v42
-; GCN-NEXT: v_add_f32_e32 v41, 0x38000000, v41
-; GCN-NEXT: v_add_f32_e32 v40, 0x38000000, v40
-; GCN-NEXT: v_add_f32_e32 v55, 0x38000000, v55
-; GCN-NEXT: v_add_f32_e32 v54, 0x38000000, v54
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v35, 0x38000000, v35
-; GCN-NEXT: v_add_f32_e32 v34, 0x38000000, v34
-; GCN-NEXT: v_add_f32_e32 v33, 0x38000000, v33
-; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v45
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v56
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v57
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v58
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v59
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v60
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v62
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v19
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v8
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v10
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v12
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v13
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v14
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v15
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v16
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v17
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v18
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v4, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v7
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v23
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v25
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v27
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v34
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v35
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v36
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v37
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v38
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v39
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v48
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v49
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v50
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v51
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v52
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v53
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v54
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v55
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v40
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v42
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v43
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v44
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v45
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v46
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v47
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v56
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v57
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v58
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v59
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v60
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v61
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v62
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB51_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v45, v1, v2, 16
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v44, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v47, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v46, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v2, v1, v2, 16
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 20, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v3, v1, v3, 16
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v5, v1, v5, 16
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v7, v1, v7, 16
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v9, v1, v9, 16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 36, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v11, v1, v11, 16
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v13, v1, v13, 16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 44, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v15, v1, v15, 16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v17, v1, v17, 16
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v19, v1, v19, 16
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v21, v1, v21, 16
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v23, v1, v23, 16
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v25, v1, v25, 16
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v27, v1, v27, 16
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v29, v1, v29, 16
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v31, v1, v31, 16
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v33, v1, v33, 16
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v35, v1, v35, 16
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v37, v1, v37, 16
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v39, v1, v39, 16
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v49, v1, v49, 16
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v51, v1, v51, 16
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v53, v1, v53, 16
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v55, v1, v55, 16
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v41, v1, v41, 16
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v45, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v56, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v19, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v24, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v25, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v30, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v51, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v53, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v55, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB51_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
-; VI-NEXT: v_add_f16_sdwa v15, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v33, v15
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v14
-; VI-NEXT: v_add_f16_sdwa v14, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v33, v14
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v13
-; VI-NEXT: v_add_f16_sdwa v13, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v33, v13
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v12
-; VI-NEXT: v_add_f16_sdwa v12, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v33, v12
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v11
-; VI-NEXT: v_add_f16_sdwa v11, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v33, v11
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v10
-; VI-NEXT: v_add_f16_sdwa v10, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v33, v10
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v9
-; VI-NEXT: v_add_f16_sdwa v9, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v33, v9
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v8, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v33, v8
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v33, v7
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v33, v6
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v33, v5
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v33, v4
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v33, v3
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v33, v2
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v33, v1
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v33, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v31
-; VI-NEXT: v_add_f16_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v33, v31
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v30
-; VI-NEXT: v_add_f16_sdwa v30, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v33, v30
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v29
-; VI-NEXT: v_add_f16_sdwa v29, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v33, v29
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v28
-; VI-NEXT: v_add_f16_sdwa v28, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v33, v28
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v27
-; VI-NEXT: v_add_f16_sdwa v27, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v33, v27
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v26
-; VI-NEXT: v_add_f16_sdwa v26, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v33, v26
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v25
-; VI-NEXT: v_add_f16_sdwa v25, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v33, v25
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v24
-; VI-NEXT: v_add_f16_sdwa v24, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v33, v24
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v23
-; VI-NEXT: v_add_f16_sdwa v23, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v33, v23
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v22
-; VI-NEXT: v_add_f16_sdwa v22, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v33, v22
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v21
-; VI-NEXT: v_add_f16_sdwa v21, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v33, v21
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v20
-; VI-NEXT: v_add_f16_sdwa v20, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v33, v20
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v19
-; VI-NEXT: v_add_f16_sdwa v19, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v33, v19
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v18
-; VI-NEXT: v_add_f16_sdwa v18, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v33, v18
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v17, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v33, v17
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB51_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB51_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB51_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB51_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB51_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64bf16_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v10
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v18
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v27
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v28
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v30
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v40
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v55
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v38
-; GCN-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(4)
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
-; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v59
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v58
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v57
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v56
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v47
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v46
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v45
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v43
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v42
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v54
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v50
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v49
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:112
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v48
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v39
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v5
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v16
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:132
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v26
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v11
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; kill: killed $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; kill: killed $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB52_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v36
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v35
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v26
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v34
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v33
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v26
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v31
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 16, v26
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v6
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v3
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v57
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v59
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v7
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; kill: killed $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: .LBB52_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB52_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v35
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v35
-; GCN-NEXT: v_alignbit_b32 v26, v27, v26, 16
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v33
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v33
-; GCN-NEXT: v_alignbit_b32 v26, v27, v26, 16
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v31
-; GCN-NEXT: v_alignbit_b32 v26, v27, v26, 16
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v9
-; GCN-NEXT: v_alignbit_b32 v10, v26, v10, 16
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v6
-; GCN-NEXT: v_alignbit_b32 v8, v10, v8, 16
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v3
-; GCN-NEXT: v_alignbit_b32 v4, v8, v4, 16
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v1
-; GCN-NEXT: v_alignbit_b32 v2, v4, v2, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v61
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_alignbit_b32 v4, v4, v8, 16
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v25
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_alignbit_b32 v8, v8, v10, 16
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v24
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v8
-; GCN-NEXT: v_alignbit_b32 v10, v10, v24, 16
-; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v10
-; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
-; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_alignbit_b32 v21, v22, v21, 16
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v37
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v19
-; GCN-NEXT: v_alignbit_b32 v21, v21, v22, 16
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v15
-; GCN-NEXT: v_alignbit_b32 v18, v21, v18, 16
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v13
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v14
-; GCN-NEXT: v_alignbit_b32 v16, v17, v16, 16
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v60
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v43
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v34
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v37
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v39
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v49
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v51
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v53
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v55
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff0000, v40
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v41, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v42, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v43, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v32
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v34
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v36
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v37
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v38
-; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v39
-; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v48
-; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v49
-; GCN-NEXT: v_add_f32_e32 v39, 0x40c00000, v50
-; GCN-NEXT: v_add_f32_e32 v56, 0x40c00000, v51
-; GCN-NEXT: v_add_f32_e32 v48, 0x40c00000, v52
-; GCN-NEXT: v_add_f32_e32 v57, 0x40c00000, v53
-; GCN-NEXT: v_add_f32_e32 v49, 0x40c00000, v54
-; GCN-NEXT: v_add_f32_e32 v59, 0x40c00000, v55
-; GCN-NEXT: v_add_f32_e32 v50, 0x40c00000, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v11
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v5
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v17
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v27
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v28
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v29
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v30
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v37
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v39
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v48
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v49
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v50
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v35
-; GCN-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v39, v51, v12, 16
-; GCN-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v49, v53, v7, 16
-; GCN-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v50, v54, v16, 16
-; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_alignbit_b32 v51, v21, v18, 16
-; GCN-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_alignbit_b32 v53, v23, v22, 16
-; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_alignbit_b32 v54, v25, v24, 16
-; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v55, v26, v41, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v26, v40
-; GCN-NEXT: v_alignbit_b32 v40, v26, v42, 16
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v42, v44, v43, 16
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_alignbit_b32 v44, v45, v32, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v45, v27, v34, 16
-; GCN-NEXT: v_alignbit_b32 v46, v28, v36, 16
-; GCN-NEXT: v_alignbit_b32 v47, v29, v38, 16
-; GCN-NEXT: v_alignbit_b32 v56, v30, v56, 16
-; GCN-NEXT: v_alignbit_b32 v58, v48, v57, 16
-; GCN-NEXT: v_alignbit_b32 v62, v52, v59, 16
-; GCN-NEXT: v_alignbit_b32 v7, v62, v20, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v58, v19, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v56, v17, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v7, v47, v9, 16
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v6, v46, v6, 16
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v3, v45, v3, 16
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v1, v44, v1, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v42, v2, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v40, v4, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v55, v8, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v10, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v53, v15, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v14, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v50, v13, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v49, v5, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v11, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB52_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v52
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v59, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v48
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_add_i32_e32 v63, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; GCN-NEXT: v_or_b32_e32 v56, v1, v2
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v29
-; GCN-NEXT: v_or_b32_e32 v2, v2, v3
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v28
-; GCN-NEXT: v_or_b32_e32 v6, v6, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v27
-; GCN-NEXT: v_or_b32_e32 v10, v10, v11
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v44
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v42
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v26
-; GCN-NEXT: v_or_b32_e32 v22, v22, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v25
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v55
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_or_b32_e32 v28, v28, v29
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v54
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GCN-NEXT: v_or_b32_e32 v30, v30, v31
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v32, v32, v33
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v53
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_or_b32_e32 v34, v34, v35
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GCN-NEXT: v_or_b32_e32 v36, v36, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff, v51
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v38, v38, v48
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff, v51
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v51, v51, v52
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_or_b32_e32 v50, v50, v53
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_or_b32_e32 v54, v54, v55
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff, v49
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v49, v49, v40
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff, v41
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v41, v41, v42
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff, v39
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_or_b32_e32 v39, v39, v43
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v59, v63, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v62, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v61, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v60, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v51, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64bf16_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB52_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; VI-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; VI-NEXT: v_bfe_u32 v33, v32, 16, 1
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v32
-; VI-NEXT: v_add_u32_e32 v33, vcc, 0x7fff, v33
-; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; VI-NEXT: v_bfe_u32 v33, v16, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v33, vcc, v33, v16
-; VI-NEXT: v_add_u32_e32 v33, vcc, s6, v33
-; VI-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; VI-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; VI-NEXT: v_bfe_u32 v34, v33, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v33
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; VI-NEXT: v_bfe_u32 v34, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v34, vcc, v34, v17
-; VI-NEXT: v_add_u32_e32 v34, vcc, s6, v34
-; VI-NEXT: v_or_b32_e32 v35, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
-; VI-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; VI-NEXT: v_bfe_u32 v35, v34, 16, 1
-; VI-NEXT: v_add_u32_e32 v35, vcc, v35, v34
-; VI-NEXT: v_add_u32_e32 v35, vcc, s6, v35
-; VI-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; VI-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v34, v34
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc
-; VI-NEXT: v_bfe_u32 v35, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v35, vcc, v35, v18
-; VI-NEXT: v_add_u32_e32 v35, vcc, s6, v35
-; VI-NEXT: v_or_b32_e32 v36, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_cndmask_b32_e32 v18, v35, v36, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
-; VI-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; VI-NEXT: v_bfe_u32 v36, v35, 16, 1
-; VI-NEXT: v_add_u32_e32 v36, vcc, v36, v35
-; VI-NEXT: v_add_u32_e32 v36, vcc, s6, v36
-; VI-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; VI-NEXT: v_or_b32_e32 v37, 0x400000, v35
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v35, v35
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_cndmask_b32_e32 v35, v36, v37, vcc
-; VI-NEXT: v_bfe_u32 v36, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v36, vcc, v36, v19
-; VI-NEXT: v_add_u32_e32 v36, vcc, s6, v36
-; VI-NEXT: v_or_b32_e32 v37, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_cndmask_b32_e32 v19, v36, v37, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v20
-; VI-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; VI-NEXT: v_bfe_u32 v37, v36, 16, 1
-; VI-NEXT: v_add_u32_e32 v37, vcc, v37, v36
-; VI-NEXT: v_add_u32_e32 v37, vcc, s6, v37
-; VI-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; VI-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v36, v36
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_cndmask_b32_e32 v36, v37, v38, vcc
-; VI-NEXT: v_bfe_u32 v37, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v37, vcc, v37, v20
-; VI-NEXT: v_add_u32_e32 v37, vcc, s6, v37
-; VI-NEXT: v_or_b32_e32 v38, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_cndmask_b32_e32 v20, v37, v38, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; VI-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
-; VI-NEXT: v_bfe_u32 v38, v37, 16, 1
-; VI-NEXT: v_add_u32_e32 v38, vcc, v38, v37
-; VI-NEXT: v_add_u32_e32 v38, vcc, s6, v38
-; VI-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; VI-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v37, v37
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_cndmask_b32_e32 v37, v38, v39, vcc
-; VI-NEXT: v_bfe_u32 v38, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v38, vcc, v38, v21
-; VI-NEXT: v_add_u32_e32 v38, vcc, s6, v38
-; VI-NEXT: v_or_b32_e32 v39, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_cndmask_b32_e32 v21, v38, v39, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v22
-; VI-NEXT: v_add_f32_e32 v38, 0x40c00000, v38
-; VI-NEXT: v_bfe_u32 v39, v38, 16, 1
-; VI-NEXT: v_add_u32_e32 v39, vcc, v39, v38
-; VI-NEXT: v_add_u32_e32 v39, vcc, s6, v39
-; VI-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; VI-NEXT: v_or_b32_e32 v48, 0x400000, v38
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v38, v38
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_cndmask_b32_e32 v38, v39, v48, vcc
-; VI-NEXT: v_bfe_u32 v39, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v39, vcc, v39, v22
-; VI-NEXT: v_add_u32_e32 v39, vcc, s6, v39
-; VI-NEXT: v_or_b32_e32 v48, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_cndmask_b32_e32 v22, v39, v48, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v23
-; VI-NEXT: v_add_f32_e32 v39, 0x40c00000, v39
-; VI-NEXT: v_bfe_u32 v48, v39, 16, 1
-; VI-NEXT: v_add_u32_e32 v48, vcc, v48, v39
-; VI-NEXT: v_add_u32_e32 v48, vcc, s6, v48
-; VI-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; VI-NEXT: v_or_b32_e32 v49, 0x400000, v39
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v39, v39
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_cndmask_b32_e32 v39, v48, v49, vcc
-; VI-NEXT: v_bfe_u32 v48, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v48, vcc, v48, v23
-; VI-NEXT: v_add_u32_e32 v48, vcc, s6, v48
-; VI-NEXT: v_or_b32_e32 v49, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v24
-; VI-NEXT: v_add_f32_e32 v48, 0x40c00000, v48
-; VI-NEXT: v_bfe_u32 v49, v48, 16, 1
-; VI-NEXT: v_add_u32_e32 v49, vcc, v49, v48
-; VI-NEXT: v_add_u32_e32 v49, vcc, s6, v49
-; VI-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; VI-NEXT: v_or_b32_e32 v50, 0x400000, v48
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v48, v48
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_cndmask_b32_e32 v48, v49, v50, vcc
-; VI-NEXT: v_bfe_u32 v49, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v49, vcc, v49, v24
-; VI-NEXT: v_add_u32_e32 v49, vcc, s6, v49
-; VI-NEXT: v_or_b32_e32 v50, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_cndmask_b32_e32 v24, v49, v50, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v25
-; VI-NEXT: v_add_f32_e32 v49, 0x40c00000, v49
-; VI-NEXT: v_bfe_u32 v50, v49, 16, 1
-; VI-NEXT: v_add_u32_e32 v50, vcc, v50, v49
-; VI-NEXT: v_add_u32_e32 v50, vcc, s6, v50
-; VI-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; VI-NEXT: v_or_b32_e32 v51, 0x400000, v49
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v49, v49
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_cndmask_b32_e32 v49, v50, v51, vcc
-; VI-NEXT: v_bfe_u32 v50, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v50, vcc, v50, v25
-; VI-NEXT: v_add_u32_e32 v50, vcc, s6, v50
-; VI-NEXT: v_or_b32_e32 v51, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_cndmask_b32_e32 v25, v50, v51, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v26
-; VI-NEXT: v_add_f32_e32 v50, 0x40c00000, v50
-; VI-NEXT: v_bfe_u32 v51, v50, 16, 1
-; VI-NEXT: v_add_u32_e32 v51, vcc, v51, v50
-; VI-NEXT: v_add_u32_e32 v51, vcc, s6, v51
-; VI-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; VI-NEXT: v_or_b32_e32 v52, 0x400000, v50
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v50, v50
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_cndmask_b32_e32 v50, v51, v52, vcc
-; VI-NEXT: v_bfe_u32 v51, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v51, vcc, v51, v26
-; VI-NEXT: v_add_u32_e32 v51, vcc, s6, v51
-; VI-NEXT: v_or_b32_e32 v52, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_cndmask_b32_e32 v26, v51, v52, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
-; VI-NEXT: v_add_f32_e32 v51, 0x40c00000, v51
-; VI-NEXT: v_bfe_u32 v52, v51, 16, 1
-; VI-NEXT: v_add_u32_e32 v52, vcc, v52, v51
-; VI-NEXT: v_add_u32_e32 v52, vcc, s6, v52
-; VI-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; VI-NEXT: v_or_b32_e32 v53, 0x400000, v51
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v51, v51
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_cndmask_b32_e32 v51, v52, v53, vcc
-; VI-NEXT: v_bfe_u32 v52, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v52, vcc, v52, v27
-; VI-NEXT: v_add_u32_e32 v52, vcc, s6, v52
-; VI-NEXT: v_or_b32_e32 v53, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_cndmask_b32_e32 v27, v52, v53, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v28
-; VI-NEXT: v_add_f32_e32 v52, 0x40c00000, v52
-; VI-NEXT: v_bfe_u32 v53, v52, 16, 1
-; VI-NEXT: v_add_u32_e32 v53, vcc, v53, v52
-; VI-NEXT: v_add_u32_e32 v53, vcc, s6, v53
-; VI-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; VI-NEXT: v_or_b32_e32 v54, 0x400000, v52
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v52, v52
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_cndmask_b32_e32 v52, v53, v54, vcc
-; VI-NEXT: v_bfe_u32 v53, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v53, vcc, v53, v28
-; VI-NEXT: v_add_u32_e32 v53, vcc, s6, v53
-; VI-NEXT: v_or_b32_e32 v54, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v29
-; VI-NEXT: v_add_f32_e32 v53, 0x40c00000, v53
-; VI-NEXT: v_bfe_u32 v54, v53, 16, 1
-; VI-NEXT: v_add_u32_e32 v54, vcc, v54, v53
-; VI-NEXT: v_add_u32_e32 v54, vcc, s6, v54
-; VI-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; VI-NEXT: v_or_b32_e32 v55, 0x400000, v53
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v53, v53
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_cndmask_b32_e32 v53, v54, v55, vcc
-; VI-NEXT: v_bfe_u32 v54, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v54, vcc, v54, v29
-; VI-NEXT: v_add_u32_e32 v54, vcc, s6, v54
-; VI-NEXT: v_or_b32_e32 v55, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_cndmask_b32_e32 v29, v54, v55, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v30
-; VI-NEXT: v_add_f32_e32 v54, 0x40c00000, v54
-; VI-NEXT: v_bfe_u32 v55, v54, 16, 1
-; VI-NEXT: v_add_u32_e32 v55, vcc, v55, v54
-; VI-NEXT: v_add_u32_e32 v55, vcc, s6, v55
-; VI-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; VI-NEXT: v_or_b32_e32 v40, 0x400000, v54
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v54, v54
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_cndmask_b32_e32 v54, v55, v40, vcc
-; VI-NEXT: v_bfe_u32 v55, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v55, vcc, v55, v30
-; VI-NEXT: v_add_u32_e32 v55, vcc, s6, v55
-; VI-NEXT: v_or_b32_e32 v40, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_cndmask_b32_e32 v30, v55, v40, vcc
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v31
-; VI-NEXT: v_add_f32_e32 v55, 0x40c00000, v55
-; VI-NEXT: v_bfe_u32 v40, v55, 16, 1
-; VI-NEXT: v_add_u32_e32 v40, vcc, v40, v55
-; VI-NEXT: v_add_u32_e32 v40, vcc, s6, v40
-; VI-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; VI-NEXT: v_or_b32_e32 v41, 0x400000, v55
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v55, v55
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_cndmask_b32_e32 v55, v40, v41, vcc
-; VI-NEXT: v_bfe_u32 v40, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v40, vcc, v40, v31
-; VI-NEXT: v_add_u32_e32 v40, vcc, s6, v40
-; VI-NEXT: v_or_b32_e32 v41, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_cndmask_b32_e32 v31, v40, v41, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v0
-; VI-NEXT: v_add_f32_e32 v40, 0x40c00000, v40
-; VI-NEXT: v_bfe_u32 v41, v40, 16, 1
-; VI-NEXT: v_add_u32_e32 v41, vcc, v41, v40
-; VI-NEXT: v_add_u32_e32 v41, vcc, s6, v41
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v42, 0x400000, v40
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v40, v40
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v40, v41, v42, vcc
-; VI-NEXT: v_bfe_u32 v41, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v41, vcc, v41, v0
-; VI-NEXT: v_add_u32_e32 v41, vcc, s6, v41
-; VI-NEXT: v_or_b32_e32 v42, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v17, v41, v42, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v1
-; VI-NEXT: v_add_f32_e32 v41, 0x40c00000, v41
-; VI-NEXT: v_bfe_u32 v42, v41, 16, 1
-; VI-NEXT: v_add_u32_e32 v42, vcc, v42, v41
-; VI-NEXT: v_add_u32_e32 v42, vcc, s6, v42
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v43, 0x400000, v41
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v41, v41
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v41, v42, v43, vcc
-; VI-NEXT: v_bfe_u32 v42, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v42, vcc, v42, v1
-; VI-NEXT: v_add_u32_e32 v42, vcc, s6, v42
-; VI-NEXT: v_or_b32_e32 v43, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_cndmask_b32_e32 v16, v42, v43, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v2
-; VI-NEXT: v_add_f32_e32 v42, 0x40c00000, v42
-; VI-NEXT: v_bfe_u32 v43, v42, 16, 1
-; VI-NEXT: v_add_u32_e32 v43, vcc, v43, v42
-; VI-NEXT: v_add_u32_e32 v43, vcc, s6, v43
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v44, 0x400000, v42
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v42, v42
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v42, v43, v44, vcc
-; VI-NEXT: v_bfe_u32 v43, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v43, vcc, v43, v2
-; VI-NEXT: v_add_u32_e32 v43, vcc, s6, v43
-; VI-NEXT: v_or_b32_e32 v44, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v43, v44, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v3
-; VI-NEXT: v_add_f32_e32 v43, 0x40c00000, v43
-; VI-NEXT: v_bfe_u32 v44, v43, 16, 1
-; VI-NEXT: v_add_u32_e32 v44, vcc, v44, v43
-; VI-NEXT: v_add_u32_e32 v44, vcc, s6, v44
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v45, 0x400000, v43
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v43, v43
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v43, v44, v45, vcc
-; VI-NEXT: v_bfe_u32 v44, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v44, vcc, v44, v3
-; VI-NEXT: v_add_u32_e32 v44, vcc, s6, v44
-; VI-NEXT: v_or_b32_e32 v45, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v44, v45, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v4
-; VI-NEXT: v_add_f32_e32 v44, 0x40c00000, v44
-; VI-NEXT: v_bfe_u32 v45, v44, 16, 1
-; VI-NEXT: v_add_u32_e32 v45, vcc, v45, v44
-; VI-NEXT: v_add_u32_e32 v45, vcc, s6, v45
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v46, 0x400000, v44
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v44, v44
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v44, v45, v46, vcc
-; VI-NEXT: v_bfe_u32 v45, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v45, vcc, v45, v4
-; VI-NEXT: v_add_u32_e32 v45, vcc, s6, v45
-; VI-NEXT: v_or_b32_e32 v46, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v45, v46, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v5
-; VI-NEXT: v_add_f32_e32 v45, 0x40c00000, v45
-; VI-NEXT: v_bfe_u32 v46, v45, 16, 1
-; VI-NEXT: v_add_u32_e32 v46, vcc, v46, v45
-; VI-NEXT: v_add_u32_e32 v46, vcc, s6, v46
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v47, 0x400000, v45
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v45, v45
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v45, v46, v47, vcc
-; VI-NEXT: v_bfe_u32 v46, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v46, vcc, v46, v5
-; VI-NEXT: v_add_u32_e32 v46, vcc, s6, v46
-; VI-NEXT: v_or_b32_e32 v47, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v46, v47, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; VI-NEXT: v_add_f32_e32 v46, 0x40c00000, v46
-; VI-NEXT: v_bfe_u32 v47, v46, 16, 1
-; VI-NEXT: v_add_u32_e32 v47, vcc, v47, v46
-; VI-NEXT: v_add_u32_e32 v47, vcc, s6, v47
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v56, 0x400000, v46
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v46, v46
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v46, v47, v56, vcc
-; VI-NEXT: v_bfe_u32 v47, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v47, vcc, v47, v6
-; VI-NEXT: v_add_u32_e32 v47, vcc, s6, v47
-; VI-NEXT: v_or_b32_e32 v56, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v47, v56, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v7
-; VI-NEXT: v_add_f32_e32 v47, 0x40c00000, v47
-; VI-NEXT: v_bfe_u32 v56, v47, 16, 1
-; VI-NEXT: v_add_u32_e32 v56, vcc, v56, v47
-; VI-NEXT: v_add_u32_e32 v56, vcc, s6, v56
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v57, 0x400000, v47
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v47, v47
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v47, v56, v57, vcc
-; VI-NEXT: v_bfe_u32 v56, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v56, vcc, v56, v7
-; VI-NEXT: v_add_u32_e32 v56, vcc, s6, v56
-; VI-NEXT: v_or_b32_e32 v57, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v56, v57, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v56, 16, v8
-; VI-NEXT: v_add_f32_e32 v56, 0x40c00000, v56
-; VI-NEXT: v_bfe_u32 v57, v56, 16, 1
-; VI-NEXT: v_add_u32_e32 v57, vcc, v57, v56
-; VI-NEXT: v_add_u32_e32 v57, vcc, s6, v57
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v58, 0x400000, v56
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v56, v56
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v56, v57, v58, vcc
-; VI-NEXT: v_bfe_u32 v57, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v57, vcc, v57, v8
-; VI-NEXT: v_add_u32_e32 v57, vcc, s6, v57
-; VI-NEXT: v_or_b32_e32 v58, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v57, v58, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v57, 16, v9
-; VI-NEXT: v_add_f32_e32 v57, 0x40c00000, v57
-; VI-NEXT: v_bfe_u32 v58, v57, 16, 1
-; VI-NEXT: v_add_u32_e32 v58, vcc, v58, v57
-; VI-NEXT: v_add_u32_e32 v58, vcc, s6, v58
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v59, 0x400000, v57
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v57, v57
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v57, v58, v59, vcc
-; VI-NEXT: v_bfe_u32 v58, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v58, vcc, v58, v9
-; VI-NEXT: v_add_u32_e32 v58, vcc, s6, v58
-; VI-NEXT: v_or_b32_e32 v59, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v58, v59, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v58, 16, v10
-; VI-NEXT: v_add_f32_e32 v58, 0x40c00000, v58
-; VI-NEXT: v_bfe_u32 v59, v58, 16, 1
-; VI-NEXT: v_add_u32_e32 v59, vcc, v59, v58
-; VI-NEXT: v_add_u32_e32 v59, vcc, s6, v59
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v60, 0x400000, v58
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v58, v58
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v58, v59, v60, vcc
-; VI-NEXT: v_bfe_u32 v59, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v59, vcc, v59, v10
-; VI-NEXT: v_add_u32_e32 v59, vcc, s6, v59
-; VI-NEXT: v_or_b32_e32 v60, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v59, v60, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v59, 16, v11
-; VI-NEXT: v_add_f32_e32 v59, 0x40c00000, v59
-; VI-NEXT: v_bfe_u32 v60, v59, 16, 1
-; VI-NEXT: v_add_u32_e32 v60, vcc, v60, v59
-; VI-NEXT: v_add_u32_e32 v60, vcc, s6, v60
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v61, 0x400000, v59
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v59, v59
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v59, v60, v61, vcc
-; VI-NEXT: v_bfe_u32 v60, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v60, vcc, v60, v11
-; VI-NEXT: v_add_u32_e32 v60, vcc, s6, v60
-; VI-NEXT: v_or_b32_e32 v61, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v60, v61, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v60, 16, v12
-; VI-NEXT: v_add_f32_e32 v60, 0x40c00000, v60
-; VI-NEXT: v_bfe_u32 v61, v60, 16, 1
-; VI-NEXT: v_add_u32_e32 v61, vcc, v61, v60
-; VI-NEXT: v_add_u32_e32 v61, vcc, s6, v61
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v62, 0x400000, v60
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v60, v60
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v60, v61, v62, vcc
-; VI-NEXT: v_bfe_u32 v61, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v61, vcc, v61, v12
-; VI-NEXT: v_add_u32_e32 v61, vcc, s6, v61
-; VI-NEXT: v_or_b32_e32 v62, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v61, v62, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v61, 16, v13
-; VI-NEXT: v_add_f32_e32 v61, 0x40c00000, v61
-; VI-NEXT: v_bfe_u32 v62, v61, 16, 1
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v61
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v61
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v61, v61
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v61, v62, v63, vcc
-; VI-NEXT: v_bfe_u32 v62, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v13
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v62, v63, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v62, 16, v14
-; VI-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; VI-NEXT: v_bfe_u32 v63, v62, 16, 1
-; VI-NEXT: v_add_u32_e32 v63, vcc, v63, v62
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_add_u32_e32 v63, vcc, s6, v63
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; VI-NEXT: v_bfe_u32 v62, v14, 16, 1
-; VI-NEXT: v_cndmask_b32_e32 v1, v63, v0, vcc
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v14
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v62, v63, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v62, 16, v15
-; VI-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; VI-NEXT: v_bfe_u32 v63, v62, 16, 1
-; VI-NEXT: v_add_u32_e32 v63, vcc, v63, v62
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_add_u32_e32 v63, vcc, s6, v63
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; VI-NEXT: v_bfe_u32 v62, v15, 16, 1
-; VI-NEXT: v_cndmask_b32_e32 v0, v63, v0, vcc
-; VI-NEXT: v_add_u32_e32 v62, vcc, v62, v15
-; VI-NEXT: v_add_u32_e32 v62, vcc, s6, v62
-; VI-NEXT: v_or_b32_e32 v63, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v62, v63, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v0, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v0, v1, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v0, v61, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v0, v60, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v0, v59, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v0, v58, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v0, v57, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v0, v56, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v0, v47, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v0, v46, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v0, v45, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v0, v44, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v0, v43, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v0, v42, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v31
-; VI-NEXT: v_alignbit_b32 v31, v16, v55, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v30
-; VI-NEXT: v_alignbit_b32 v30, v16, v54, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v29
-; VI-NEXT: v_alignbit_b32 v29, v16, v53, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v28
-; VI-NEXT: v_alignbit_b32 v28, v16, v52, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v27
-; VI-NEXT: v_alignbit_b32 v27, v16, v51, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v26
-; VI-NEXT: v_alignbit_b32 v26, v16, v50, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v25
-; VI-NEXT: v_alignbit_b32 v25, v16, v49, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v24
-; VI-NEXT: v_alignbit_b32 v24, v16, v48, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v23
-; VI-NEXT: v_alignbit_b32 v23, v16, v39, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v22
-; VI-NEXT: v_alignbit_b32 v22, v16, v38, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v21
-; VI-NEXT: v_alignbit_b32 v21, v16, v37, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v20
-; VI-NEXT: v_alignbit_b32 v20, v16, v36, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v19
-; VI-NEXT: v_alignbit_b32 v19, v16, v35, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v18
-; VI-NEXT: v_alignbit_b32 v18, v16, v34, 16
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v32
-; VI-NEXT: v_alignbit_b32 v1, v0, v41, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v17
-; VI-NEXT: v_alignbit_b32 v17, v16, v33, 16
-; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: v_alignbit_b32 v0, v0, v40, 16
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_alignbit_b32 v16, v16, v32, 16
-; VI-NEXT: .LBB52_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64bf16_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB52_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX9-NEXT: v_add3_u32 v33, v33, v32, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
-; GFX9-NEXT: v_bfe_u32 v33, v16, 16, 1
-; GFX9-NEXT: v_add3_u32 v33, v33, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v33, 0x40c00000, v33
-; GFX9-NEXT: v_bfe_u32 v34, v33, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX9-NEXT: v_add3_u32 v34, v34, v33, s6
-; GFX9-NEXT: v_or_b32_e32 v35, 0x400000, v33
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_cndmask_b32_e32 v33, v34, v35, vcc
-; GFX9-NEXT: v_bfe_u32 v34, v17, 16, 1
-; GFX9-NEXT: v_add3_u32 v34, v34, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v35, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v32, v34, v35, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX9-NEXT: v_bfe_u32 v35, v34, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX9-NEXT: v_add3_u32 v35, v35, v34, s6
-; GFX9-NEXT: v_or_b32_e32 v36, 0x400000, v34
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v34, v34
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v34, v35, v36, vcc
-; GFX9-NEXT: v_bfe_u32 v35, v18, 16, 1
-; GFX9-NEXT: v_add3_u32 v35, v35, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v36, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v35, v36, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v35, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX9-NEXT: v_bfe_u32 v36, v35, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX9-NEXT: v_add3_u32 v36, v36, v35, s6
-; GFX9-NEXT: v_or_b32_e32 v37, 0x400000, v35
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v35, v35
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v35, v36, v37, vcc
-; GFX9-NEXT: v_bfe_u32 v36, v19, 16, 1
-; GFX9-NEXT: v_add3_u32 v36, v36, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v37, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v36, v37, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v36, 16, v20
-; GFX9-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX9-NEXT: v_bfe_u32 v37, v36, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX9-NEXT: v_add3_u32 v37, v37, v36, s6
-; GFX9-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v36, v36
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v36, v37, v38, vcc
-; GFX9-NEXT: v_bfe_u32 v37, v20, 16, 1
-; GFX9-NEXT: v_add3_u32 v37, v37, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v38, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v37, v38, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v37, 16, v21
-; GFX9-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
-; GFX9-NEXT: v_bfe_u32 v38, v37, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX9-NEXT: v_add3_u32 v38, v38, v37, s6
-; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v37, v37
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v37, v38, v39, vcc
-; GFX9-NEXT: v_bfe_u32 v38, v21, 16, 1
-; GFX9-NEXT: v_add3_u32 v38, v38, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v38, v39, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v38, 16, v22
-; GFX9-NEXT: v_add_f32_e32 v38, 0x40c00000, v38
-; GFX9-NEXT: v_bfe_u32 v39, v38, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX9-NEXT: v_add3_u32 v39, v39, v38, s6
-; GFX9-NEXT: v_or_b32_e32 v48, 0x400000, v38
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v38, v38
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v38, v39, v48, vcc
-; GFX9-NEXT: v_bfe_u32 v39, v22, 16, 1
-; GFX9-NEXT: v_add3_u32 v39, v39, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v48, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v39, v48, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v39, 16, v23
-; GFX9-NEXT: v_add_f32_e32 v39, 0x40c00000, v39
-; GFX9-NEXT: v_bfe_u32 v48, v39, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX9-NEXT: v_add3_u32 v48, v48, v39, s6
-; GFX9-NEXT: v_or_b32_e32 v49, 0x400000, v39
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v39, v39
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v39, v48, v49, vcc
-; GFX9-NEXT: v_bfe_u32 v48, v23, 16, 1
-; GFX9-NEXT: v_add3_u32 v48, v48, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v49, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v48, 16, v24
-; GFX9-NEXT: v_add_f32_e32 v48, 0x40c00000, v48
-; GFX9-NEXT: v_bfe_u32 v49, v48, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX9-NEXT: v_add3_u32 v49, v49, v48, s6
-; GFX9-NEXT: v_or_b32_e32 v50, 0x400000, v48
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v48, v48
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v48, v49, v50, vcc
-; GFX9-NEXT: v_bfe_u32 v49, v24, 16, 1
-; GFX9-NEXT: v_add3_u32 v49, v49, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v50, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v49, v50, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v49, 16, v25
-; GFX9-NEXT: v_add_f32_e32 v49, 0x40c00000, v49
-; GFX9-NEXT: v_bfe_u32 v50, v49, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX9-NEXT: v_add3_u32 v50, v50, v49, s6
-; GFX9-NEXT: v_or_b32_e32 v51, 0x400000, v49
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v49, v49
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v49, v50, v51, vcc
-; GFX9-NEXT: v_bfe_u32 v50, v25, 16, 1
-; GFX9-NEXT: v_add3_u32 v50, v50, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v51, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v50, v51, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v50, 16, v26
-; GFX9-NEXT: v_add_f32_e32 v50, 0x40c00000, v50
-; GFX9-NEXT: v_bfe_u32 v51, v50, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX9-NEXT: v_add3_u32 v51, v51, v50, s6
-; GFX9-NEXT: v_or_b32_e32 v52, 0x400000, v50
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v50, v50
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v50, v51, v52, vcc
-; GFX9-NEXT: v_bfe_u32 v51, v26, 16, 1
-; GFX9-NEXT: v_add3_u32 v51, v51, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v52, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v51, v52, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v51, 16, v27
-; GFX9-NEXT: v_add_f32_e32 v51, 0x40c00000, v51
-; GFX9-NEXT: v_bfe_u32 v52, v51, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX9-NEXT: v_add3_u32 v52, v52, v51, s6
-; GFX9-NEXT: v_or_b32_e32 v53, 0x400000, v51
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v51, v51
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v51, v52, v53, vcc
-; GFX9-NEXT: v_bfe_u32 v52, v27, 16, 1
-; GFX9-NEXT: v_add3_u32 v52, v52, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v53, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v52, v53, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v52, 16, v28
-; GFX9-NEXT: v_add_f32_e32 v52, 0x40c00000, v52
-; GFX9-NEXT: v_bfe_u32 v53, v52, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX9-NEXT: v_add3_u32 v53, v53, v52, s6
-; GFX9-NEXT: v_or_b32_e32 v54, 0x400000, v52
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v52, v52
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v52, v53, v54, vcc
-; GFX9-NEXT: v_bfe_u32 v53, v28, 16, 1
-; GFX9-NEXT: v_add3_u32 v53, v53, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v54, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v53, 16, v29
-; GFX9-NEXT: v_add_f32_e32 v53, 0x40c00000, v53
-; GFX9-NEXT: v_bfe_u32 v54, v53, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX9-NEXT: v_add3_u32 v54, v54, v53, s6
-; GFX9-NEXT: v_or_b32_e32 v55, 0x400000, v53
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v53, v53
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v53, v54, v55, vcc
-; GFX9-NEXT: v_bfe_u32 v54, v29, 16, 1
-; GFX9-NEXT: v_add3_u32 v54, v54, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v55, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v54, v55, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v54, 16, v30
-; GFX9-NEXT: v_add_f32_e32 v54, 0x40c00000, v54
-; GFX9-NEXT: v_bfe_u32 v55, v54, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX9-NEXT: v_add3_u32 v55, v55, v54, s6
-; GFX9-NEXT: v_or_b32_e32 v40, 0x400000, v54
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v54, v54
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v54, v55, v40, vcc
-; GFX9-NEXT: v_bfe_u32 v55, v30, 16, 1
-; GFX9-NEXT: v_add3_u32 v55, v55, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v40, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v55, v40, vcc
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v55, 16, v31
-; GFX9-NEXT: v_add_f32_e32 v55, 0x40c00000, v55
-; GFX9-NEXT: v_bfe_u32 v40, v55, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX9-NEXT: v_add3_u32 v40, v40, v55, s6
-; GFX9-NEXT: v_or_b32_e32 v41, 0x400000, v55
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v55, v55
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v55, v40, v41, vcc
-; GFX9-NEXT: v_bfe_u32 v40, v31, 16, 1
-; GFX9-NEXT: v_add3_u32 v40, v40, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v41, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v40, v41, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v40, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v40, 0x40c00000, v40
-; GFX9-NEXT: v_bfe_u32 v41, v40, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v41, v41, v40, s6
-; GFX9-NEXT: v_or_b32_e32 v42, 0x400000, v40
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v40, v40
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v40, v41, v42, vcc
-; GFX9-NEXT: v_bfe_u32 v41, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v41, v41, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v42, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v41, v42, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v41, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v41, 0x40c00000, v41
-; GFX9-NEXT: v_bfe_u32 v42, v41, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v42, v42, v41, s6
-; GFX9-NEXT: v_or_b32_e32 v43, 0x400000, v41
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v41, v41
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v41, v42, v43, vcc
-; GFX9-NEXT: v_bfe_u32 v42, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v42, v42, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v43, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v42, v43, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v42, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v42, 0x40c00000, v42
-; GFX9-NEXT: v_bfe_u32 v43, v42, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v43, v43, v42, s6
-; GFX9-NEXT: v_or_b32_e32 v44, 0x400000, v42
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v42, v42
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v42, v43, v44, vcc
-; GFX9-NEXT: v_bfe_u32 v43, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v43, v43, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v44, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v43, v44, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v43, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v43, 0x40c00000, v43
-; GFX9-NEXT: v_bfe_u32 v44, v43, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v44, v44, v43, s6
-; GFX9-NEXT: v_or_b32_e32 v45, 0x400000, v43
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v43, v43
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v43, v44, v45, vcc
-; GFX9-NEXT: v_bfe_u32 v44, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v44, v44, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v45, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v44, v45, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v44, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v44, 0x40c00000, v44
-; GFX9-NEXT: v_bfe_u32 v45, v44, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v45, v45, v44, s6
-; GFX9-NEXT: v_or_b32_e32 v46, 0x400000, v44
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v44, v44
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v44, v45, v46, vcc
-; GFX9-NEXT: v_bfe_u32 v45, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v45, v45, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v46, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v45, v46, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v45, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v45, 0x40c00000, v45
-; GFX9-NEXT: v_bfe_u32 v46, v45, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v46, v46, v45, s6
-; GFX9-NEXT: v_or_b32_e32 v47, 0x400000, v45
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v45, v45
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v45, v46, v47, vcc
-; GFX9-NEXT: v_bfe_u32 v46, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v46, v46, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v47, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v46, v47, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v46, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v46, 0x40c00000, v46
-; GFX9-NEXT: v_bfe_u32 v47, v46, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v47, v47, v46, s6
-; GFX9-NEXT: v_or_b32_e32 v56, 0x400000, v46
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v46, v46
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v46, v47, v56, vcc
-; GFX9-NEXT: v_bfe_u32 v47, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v47, v47, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v56, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v47, v56, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v47, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v47, 0x40c00000, v47
-; GFX9-NEXT: v_bfe_u32 v56, v47, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v56, v56, v47, s6
-; GFX9-NEXT: v_or_b32_e32 v57, 0x400000, v47
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v47, v47
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v47, v56, v57, vcc
-; GFX9-NEXT: v_bfe_u32 v56, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v56, v56, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v57, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v56, v57, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v56, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v56, 0x40c00000, v56
-; GFX9-NEXT: v_bfe_u32 v57, v56, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v57, v57, v56, s6
-; GFX9-NEXT: v_or_b32_e32 v58, 0x400000, v56
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v56, v56
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v56, v57, v58, vcc
-; GFX9-NEXT: v_bfe_u32 v57, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v57, v57, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v58, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v57, v58, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v57, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v57, 0x40c00000, v57
-; GFX9-NEXT: v_bfe_u32 v58, v57, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v58, v58, v57, s6
-; GFX9-NEXT: v_or_b32_e32 v59, 0x400000, v57
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v57, v57
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v57, v58, v59, vcc
-; GFX9-NEXT: v_bfe_u32 v58, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v58, v58, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v59, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v58, v59, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v58, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v58, 0x40c00000, v58
-; GFX9-NEXT: v_bfe_u32 v59, v58, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v59, v59, v58, s6
-; GFX9-NEXT: v_or_b32_e32 v60, 0x400000, v58
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v58, v58
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v58, v59, v60, vcc
-; GFX9-NEXT: v_bfe_u32 v59, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v59, v59, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v60, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v59, v60, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v59, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v59, 0x40c00000, v59
-; GFX9-NEXT: v_bfe_u32 v60, v59, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v60, v60, v59, s6
-; GFX9-NEXT: v_or_b32_e32 v61, 0x400000, v59
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v59, v59
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v59, v60, v61, vcc
-; GFX9-NEXT: v_bfe_u32 v60, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v60, v60, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v61, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v60, v61, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v60, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v60, 0x40c00000, v60
-; GFX9-NEXT: v_bfe_u32 v61, v60, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v61, v61, v60, s6
-; GFX9-NEXT: v_or_b32_e32 v62, 0x400000, v60
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v60, v60
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v60, v61, v62, vcc
-; GFX9-NEXT: v_bfe_u32 v61, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v61, v61, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v62, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v61, v62, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v61, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v61, 0x40c00000, v61
-; GFX9-NEXT: v_bfe_u32 v62, v61, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v62, v62, v61, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v61
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v61, v61
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v61, v62, v63, vcc
-; GFX9-NEXT: v_bfe_u32 v62, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v62, v62, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v62, v63, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v62, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_bfe_u32 v63, v62, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_add3_u32 v63, v63, v62, s6
-; GFX9-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; GFX9-NEXT: v_bfe_u32 v62, v14, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v63, v0, vcc
-; GFX9-NEXT: v_add3_u32 v62, v62, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v62, v63, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v62, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v62, 0x40c00000, v62
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_bfe_u32 v63, v62, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_add3_u32 v63, v63, v62, s6
-; GFX9-NEXT: v_or_b32_e32 v0, 0x400000, v62
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v62, v62
-; GFX9-NEXT: v_bfe_u32 v62, v15, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v63, v0, vcc
-; GFX9-NEXT: v_add3_u32 v62, v62, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v63, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v62, v63, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v0, s6
-; GFX9-NEXT: v_perm_b32 v14, v14, v1, s6
-; GFX9-NEXT: v_perm_b32 v1, v16, v41, s6
-; GFX9-NEXT: v_perm_b32 v0, v17, v40, s6
-; GFX9-NEXT: v_perm_b32 v17, v32, v33, s6
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v13, v13, v61, s6
-; GFX9-NEXT: v_perm_b32 v12, v12, v60, s6
-; GFX9-NEXT: v_perm_b32 v11, v11, v59, s6
-; GFX9-NEXT: v_perm_b32 v10, v10, v58, s6
-; GFX9-NEXT: v_perm_b32 v9, v9, v57, s6
-; GFX9-NEXT: v_perm_b32 v8, v8, v56, s6
-; GFX9-NEXT: v_perm_b32 v7, v7, v47, s6
-; GFX9-NEXT: v_perm_b32 v6, v6, v46, s6
-; GFX9-NEXT: v_perm_b32 v5, v5, v45, s6
-; GFX9-NEXT: v_perm_b32 v4, v4, v44, s6
-; GFX9-NEXT: v_perm_b32 v3, v3, v43, s6
-; GFX9-NEXT: v_perm_b32 v2, v2, v42, s6
-; GFX9-NEXT: v_perm_b32 v31, v31, v55, s6
-; GFX9-NEXT: v_perm_b32 v30, v30, v54, s6
-; GFX9-NEXT: v_perm_b32 v29, v29, v53, s6
-; GFX9-NEXT: v_perm_b32 v28, v28, v52, s6
-; GFX9-NEXT: v_perm_b32 v27, v27, v51, s6
-; GFX9-NEXT: v_perm_b32 v26, v26, v50, s6
-; GFX9-NEXT: v_perm_b32 v25, v25, v49, s6
-; GFX9-NEXT: v_perm_b32 v24, v24, v48, s6
-; GFX9-NEXT: v_perm_b32 v23, v23, v39, s6
-; GFX9-NEXT: v_perm_b32 v22, v22, v38, s6
-; GFX9-NEXT: v_perm_b32 v21, v21, v37, s6
-; GFX9-NEXT: v_perm_b32 v20, v20, v36, s6
-; GFX9-NEXT: v_perm_b32 v19, v19, v35, s6
-; GFX9-NEXT: v_perm_b32 v18, v18, v34, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v16, v32, v16, s6
-; GFX9-NEXT: .LBB52_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64bf16_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB52_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v32, 16, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
-; GFX11-NEXT: v_dual_add_f32 v69, 0x40c00000, v69 :: v_dual_lshlrev_b32 v70, 16, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v37, v33, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v32
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v10
-; GFX11-NEXT: v_add3_u32 v37, v37, v33, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v87, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v16
-; GFX11-NEXT: v_bfe_u32 v16, v32, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v87, 0x40c00000, v87
-; GFX11-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v34
-; GFX11-NEXT: v_add3_u32 v16, v16, v32, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v9
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v38, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v36, 0x40c00000, v17
-; GFX11-NEXT: v_bfe_u32 v17, v34, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add_f32_e32 v38, 0x40c00000, v18
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v11
-; GFX11-NEXT: v_add3_u32 v17, v17, v34, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v34, v36, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v32, v17, v39, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v33
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-NEXT: v_add3_u32 v33, v34, v36, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v37, v17, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v37, v35, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v35
-; GFX11-NEXT: v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v39, 0x40c00000, v18
-; GFX11-NEXT: v_add3_u32 v34, v37, v35, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v37, v38, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GFX11-NEXT: v_perm_b32 v17, v33, v17, 0x7060302
-; GFX11-NEXT: v_perm_b32 v16, v32, v16, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v34, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v34, v37, v38, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v36, v39, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v37, 0x40c00000, v19
-; GFX11-NEXT: v_or_b32_e32 v35, 0x400000, v38
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GFX11-NEXT: v_bfe_u32 v38, v37, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v34, v35, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v36, v39, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v39
-; GFX11-NEXT: v_add_f32_e32 v48, 0x40c00000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_add_f32_e32 v39, 0x40c00000, v20
-; GFX11-NEXT: v_perm_b32 v18, v34, v18, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v35, v38, v37, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v37
-; GFX11-NEXT: v_bfe_u32 v38, v48, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v48
-; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v35, v35, v36, vcc_lo
-; GFX11-NEXT: v_add3_u32 v36, v38, v48, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v38, v39, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_dual_add_f32 v49, 0x40c00000, v20 :: v_dual_add_f32 v48, 0x40c00000, v21
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v36, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v36, v38, v39, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v39
-; GFX11-NEXT: v_bfe_u32 v38, v49, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-NEXT: v_bfe_u32 v39, v48, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v21
-; GFX11-NEXT: v_perm_b32 v19, v35, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v36, v36, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v38, v49, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v49
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_dual_add_f32 v49, 0x40c00000, v22 :: v_dual_lshlrev_b32 v22, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GFX11-NEXT: v_perm_b32 v20, v36, v20, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v37, v38, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v39, v48, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v48
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-NEXT: v_bfe_u32 v48, v49, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v51, 0x40c00000, v22
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v37, v37, v38, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v39, v50, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_perm_b32 v21, v37, v21, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v38, v39, v50, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v50
-; GFX11-NEXT: v_add_f32_e32 v50, 0x40c00000, v23
-; GFX11-NEXT: v_cndmask_b32_e32 v22, v38, v39, vcc_lo
-; GFX11-NEXT: v_add3_u32 v38, v48, v49, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v39, 0x400000, v49
-; GFX11-NEXT: v_bfe_u32 v48, v51, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v49, v49
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v24
-; GFX11-NEXT: v_bfe_u32 v49, v50, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
-; GFX11-NEXT: v_cndmask_b32_e32 v38, v38, v39, vcc_lo
-; GFX11-NEXT: v_add3_u32 v39, v48, v51, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v51
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: v_dual_add_f32 v52, 0x40c00000, v23 :: v_dual_add_f32 v51, 0x40c00000, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v39, v48, vcc_lo
-; GFX11-NEXT: v_add3_u32 v39, v49, v50, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v50
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-NEXT: v_bfe_u32 v50, v51, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v53, 0x40c00000, v24
-; GFX11-NEXT: v_perm_b32 v22, v38, v22, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v39, v39, v48, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v49, v52, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v23, v39, v23, 0x7060302
-; GFX11-NEXT: v_add3_u32 v48, v49, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v52
-; GFX11-NEXT: v_add_f32_e32 v52, 0x40c00000, v25
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v48, v49, vcc_lo
-; GFX11-NEXT: v_add3_u32 v48, v50, v51, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v51
-; GFX11-NEXT: v_bfe_u32 v50, v53, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v26
-; GFX11-NEXT: v_bfe_u32 v51, v52, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
-; GFX11-NEXT: v_cndmask_b32_e32 v48, v48, v49, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v50, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v53
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: v_dual_add_f32 v54, 0x40c00000, v25 :: v_dual_add_f32 v53, 0x40c00000, v26
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v27
-; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v49, v50, vcc_lo
-; GFX11-NEXT: v_add3_u32 v49, v51, v52, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v52
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-NEXT: v_bfe_u32 v52, v53, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v55, 0x40c00000, v26
-; GFX11-NEXT: v_perm_b32 v24, v48, v24, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v49, v49, v50, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v51, v54, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v54, v54
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v25, v49, v25, 0x7060302
-; GFX11-NEXT: v_add3_u32 v50, v51, v54, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v54
-; GFX11-NEXT: v_add_f32_e32 v54, 0x40c00000, v27
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v26, v50, v51, vcc_lo
-; GFX11-NEXT: v_add3_u32 v50, v52, v53, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v51, 0x400000, v53
-; GFX11-NEXT: v_bfe_u32 v52, v55, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v53, v53
-; GFX11-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GFX11-NEXT: v_bfe_u32 v53, v54, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
-; GFX11-NEXT: v_cndmask_b32_e32 v50, v50, v51, vcc_lo
-; GFX11-NEXT: v_add3_u32 v51, v52, v55, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v55
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55
-; GFX11-NEXT: v_dual_add_f32 v64, 0x40c00000, v27 :: v_dual_add_f32 v55, 0x40c00000, v28
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v29
-; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
-; GFX11-NEXT: v_cndmask_b32_e32 v27, v51, v52, vcc_lo
-; GFX11-NEXT: v_add3_u32 v51, v53, v54, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v54
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v54, v54
-; GFX11-NEXT: v_bfe_u32 v54, v55, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v65, 0x40c00000, v28
-; GFX11-NEXT: v_perm_b32 v26, v50, v26, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v51, v51, v52, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v53, v64, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v64, v64
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v27, v51, v27, 0x7060302
-; GFX11-NEXT: v_add3_u32 v52, v53, v64, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v64
-; GFX11-NEXT: v_add_f32_e32 v64, 0x40c00000, v29
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v28, v52, v53, vcc_lo
-; GFX11-NEXT: v_add3_u32 v52, v54, v55, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v53, 0x400000, v55
-; GFX11-NEXT: v_bfe_u32 v54, v65, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v30
-; GFX11-NEXT: v_bfe_u32 v55, v64, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
-; GFX11-NEXT: v_cndmask_b32_e32 v52, v52, v53, vcc_lo
-; GFX11-NEXT: v_add3_u32 v53, v54, v65, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v65
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v65, v65
-; GFX11-NEXT: v_dual_add_f32 v66, 0x40c00000, v29 :: v_dual_add_f32 v65, 0x40c00000, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v31
-; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
-; GFX11-NEXT: v_cndmask_b32_e32 v29, v53, v54, vcc_lo
-; GFX11-NEXT: v_add3_u32 v53, v55, v64, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v64
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v64, v64
-; GFX11-NEXT: v_bfe_u32 v64, v65, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v67, 0x40c00000, v30
-; GFX11-NEXT: v_perm_b32 v28, v52, v28, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v53, v53, v54, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v55, v66, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v66, v66
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v29, v53, v29, 0x7060302
-; GFX11-NEXT: v_add3_u32 v54, v55, v66, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v66
-; GFX11-NEXT: v_add_f32_e32 v66, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v30, v54, v55, vcc_lo
-; GFX11-NEXT: v_add3_u32 v54, v64, v65, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v55, 0x400000, v65
-; GFX11-NEXT: v_bfe_u32 v64, v67, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v65, v65
-; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v0
-; GFX11-NEXT: v_bfe_u32 v65, v66, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v54, v54, v55, vcc_lo
-; GFX11-NEXT: v_add3_u32 v55, v64, v67, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v67
-; GFX11-NEXT: v_add_f32_e32 v68, 0x40c00000, v31
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v67, v67
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v67, 16, v1
-; GFX11-NEXT: v_perm_b32 v30, v54, v30, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v55, v64, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v67, 0x40c00000, v67
-; GFX11-NEXT: v_add3_u32 v55, v65, v66, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v66
-; GFX11-NEXT: v_bfe_u32 v65, v68, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v66, v66
-; GFX11-NEXT: v_bfe_u32 v66, v0, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v55, v55, v64, vcc_lo
-; GFX11-NEXT: v_add3_u32 v64, v65, v68, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v65, 0x400000, v68
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v68, v68
-; GFX11-NEXT: v_bfe_u32 v68, v67, 16, 1
-; GFX11-NEXT: v_perm_b32 v31, v55, v31, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v64, v64, v65, vcc_lo
-; GFX11-NEXT: v_add3_u32 v65, v66, v0, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_dual_cndmask_b32 v0, v65, v66 :: v_dual_and_b32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_add3_u32 v65, v68, v67, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v67
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v67, v67
-; GFX11-NEXT: v_perm_b32 v0, v0, v64, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v68, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v1
-; GFX11-NEXT: v_cndmask_b32_e32 v65, v65, v66, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v66, v68, v1, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v68, v69, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v66, v67, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v66, v68, v69, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v67, 0x400000, v69
-; GFX11-NEXT: v_bfe_u32 v68, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v69, v69
-; GFX11-NEXT: v_perm_b32 v1, v1, v65, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v66, v66, v67, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v67, v68, v2, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v67, v68 :: v_dual_and_b32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v70, 0x40c00000, v70 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v2, v2, v66, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v69, v70, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v70
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v70, v70
-; GFX11-NEXT: v_add3_u32 v67, v69, v70, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v69, v3, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v67, v67, v68, vcc_lo
-; GFX11-NEXT: v_add3_u32 v68, v69, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v69, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v68, v69 :: v_dual_and_b32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_dual_add_f32 v71, 0x40c00000, v71 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v3, v3, v67, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v70, v71, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v69, 0x400000, v71
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v71, v71
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v68, v70, v71, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v70, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v68, v68, v69, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v69, v70, v4, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v69, v70 :: v_dual_and_b32 v5, 0xffff0000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_perm_b32 v4, v4, v68, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v71, v80, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v80
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v80, v80
-; GFX11-NEXT: v_add3_u32 v69, v71, v80, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v71, v5, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v69, v69, v70, vcc_lo
-; GFX11-NEXT: v_add3_u32 v70, v71, v5, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v71, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v70, v71 :: v_dual_and_b32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_dual_add_f32 v81, 0x40c00000, v81 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v5, v5, v69, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v80, v81, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v71, 0x400000, v81
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v81, v81
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v70, v80, v81, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v80, v6, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v70, v70, v71, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v71, v80, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v71, v80 :: v_dual_and_b32 v7, 0xffff0000, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_perm_b32 v6, v6, v70, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v81, v82, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v82
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-NEXT: v_add3_u32 v71, v81, v82, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v81, v7, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v71, v71, v80, vcc_lo
-; GFX11-NEXT: v_add3_u32 v80, v81, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v81, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v80, v81 :: v_dual_and_b32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_dual_add_f32 v83, 0x40c00000, v83 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v7, v7, v71, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v82, v83, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v81, 0x400000, v83
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v80, v82, v83, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v82, v8, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v80, v80, v81, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v81, v82, v8, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_dual_cndmask_b32 v8, v81, v82 :: v_dual_and_b32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v84, 0x40c00000, v84 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_perm_b32 v8, v8, v80, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v83, v84, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v84
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v84, v84
-; GFX11-NEXT: v_add3_u32 v81, v83, v84, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v83, v9, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v81, v81, v82, vcc_lo
-; GFX11-NEXT: v_add3_u32 v82, v83, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v83, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v82, v83 :: v_dual_and_b32 v10, 0xffff0000, v10
-; GFX11-NEXT: v_dual_add_f32 v85, 0x40c00000, v85 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v9, v9, v81, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v84, v85, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v83, 0x400000, v85
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v85, v85
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v82, v84, v85, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v84, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v82, v82, v83, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v83, v84, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v83, v84 :: v_dual_and_b32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v86, 0x40c00000, v86 :: v_dual_add_f32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_perm_b32 v10, v10, v82, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v85, v86, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v86
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v86, v86
-; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v11
-; GFX11-NEXT: v_add3_u32 v83, v85, v86, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v85, v11, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v83, v83, v84 :: v_dual_lshlrev_b32 v84, 16, v13
-; GFX11-NEXT: v_bfe_u32 v86, v87, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_or_b32_e32 v97, 0x400000, v87
-; GFX11-NEXT: v_add_f32_e32 v84, 0x40c00000, v84
-; GFX11-NEXT: v_add3_u32 v86, v86, v87, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v87, v87
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_bfe_u32 v98, v12, 16, 1
-; GFX11-NEXT: v_bfe_u32 v99, v84, 16, 1
-; GFX11-NEXT: v_add3_u32 v85, v85, v11, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v86, v86, v97, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v97, 0x400000, v12
-; GFX11-NEXT: v_add3_u32 v87, v98, v12, 0x7fff
-; GFX11-NEXT: v_add3_u32 v98, v99, v84, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v99, 16, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v84
-; GFX11-NEXT: v_bfe_u32 v101, v13, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v12, v87, v97 :: v_dual_add_f32 v87, 0x40c00000, v99
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v84, v84
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v97, v101, v13, 0x7fff
-; GFX11-NEXT: v_perm_b32 v12, v12, v86, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v99, v87, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v84, v98, v100, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v98, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v101, 0x400000, v87
-; GFX11-NEXT: v_bfe_u32 v102, v14, 16, 1
-; GFX11-NEXT: v_add3_u32 v99, v99, v87, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v87, v87
-; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v98, 0x40c00000, v98 :: v_dual_cndmask_b32 v87, v99, v101
-; GFX11-NEXT: v_add3_u32 v101, v102, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_bfe_u32 v103, v98, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v98
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v101, v102 :: v_dual_add_f32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v103, v103, v98, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v98, v98
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v14, v14, v87, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v99, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v113, 0x400000, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v98, v103, v112, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v99, v99, v15, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v99, v113, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_perm_b32 v15, v15, v98, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v97, v100, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_perm_b32 v13, v13, v84, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-NEXT: .LBB52_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <64 x bfloat> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x bfloat> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v64bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:32
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v8
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v12
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v14
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v18
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v24
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v26
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:128
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v38
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(4)
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v55
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v54
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v35
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:116
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112
-; GCN-NEXT: s_waitcnt vmcnt(4)
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v18
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v16
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; kill: killed $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v44
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v42
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v40
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v33
-; GCN-NEXT: s_waitcnt vmcnt(12)
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v10
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: .LBB53_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v55, v10
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v14, v54, v14
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v12, v53, v12
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v16, v52, v16
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v44
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v18, v51, v18
-; GCN-NEXT: v_or_b32_e32 v20, v50, v20
-; GCN-NEXT: v_or_b32_e32 v8, v49, v8
-; GCN-NEXT: v_or_b32_e32 v22, v48, v22
-; GCN-NEXT: v_or_b32_e32 v24, v39, v24
-; GCN-NEXT: v_or_b32_e32 v26, v38, v26
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v30, v31
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v32, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v32, v32, v33
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v33, v29
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v33, v27
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v33, v25
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v33, v23
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v33, v21
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v33, v19
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v33, v17
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v33, v15
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v33, v13
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v33, v11
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v33, v9
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v33, v7
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v33, v5
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v33, v3
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v33, v1
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v10
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v33, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v34, vcc, s6, v20
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v22
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v24
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v28
-; GCN-NEXT: v_add_i32_e32 v30, vcc, s6, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, s6, v31
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v32
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v29
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v27
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v25
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v1
-; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v7
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v19
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v21
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v23
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v25
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v27
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v31
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v28
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v26
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v39, 0xffff0000, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v22
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v49, 0xffff0000, v8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v34
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v33
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v14
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v10
-; GCN-NEXT: .LBB53_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v44, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v15, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 12, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v11, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 20, v0
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v3, v2, v3, 16
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v7, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v45, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v46, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 36, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v47, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v56, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 44, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v57, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v58, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v34, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v33, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v32, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v31, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v41, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v2, v2, v4, 16
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x4c, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v30
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_alignbit_b32 v4, v4, v6, 16
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x50, v0
-; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v38
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_alignbit_b32 v6, v6, v8, 16
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 0x54, v0
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v39
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_alignbit_b32 v8, v8, v10, 16
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x58, v0
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v48
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_alignbit_b32 v10, v10, v12, 16
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 0x5c, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v49
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_alignbit_b32 v12, v12, v14, 16
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x60, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v50
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_alignbit_b32 v14, v14, v50, 16
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 0x64, v0
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_alignbit_b32 v16, v51, v16, 16
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_alignbit_b32 v22, v52, v22, 16
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 0x6c, v0
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_alignbit_b32 v24, v53, v24, 16
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x70, v0
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_alignbit_b32 v26, v54, v26, 16
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x74, v0
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_alignbit_b32 v28, v55, v28, 16
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v18, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v1, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v56, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v41, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v30, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v64bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB53_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 3
-; VI-NEXT: v_add_u16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_u16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_u16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_u16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_u16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_u16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 3, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_u16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 3, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_u16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_u16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_u16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_u16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_u16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_u16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_u16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_u16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_u16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v31, 3, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_u16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_u16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_u16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_u16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_u16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_u16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_u16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_u16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_u16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_u16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_u16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_u16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_u16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_u16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB53_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v64bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB53_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB53_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v64bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB53_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB53_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <64 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <64 x bfloat>
- br label %end
-
-end:
- %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x bfloat> %phi
-}
-
-define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64f16_to_v64i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:64
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:48
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v12
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v19
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v23
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v28
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v29
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v45
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:120
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v53
-; GCN-NEXT: s_waitcnt vmcnt(2) expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v15
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v42
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v60
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v54
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v52
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v51
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v59
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v40
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:112
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v55
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v9
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v29
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:132
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v30
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v29
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v42, v56
-; GCN-NEXT: v_mov_b32_e32 v49, v57
-; GCN-NEXT: v_mov_b32_e32 v54, v58
-; GCN-NEXT: v_mov_b32_e32 v51, v62
-; GCN-NEXT: v_mov_b32_e32 v48, v4
-; GCN-NEXT: v_mov_b32_e32 v36, v5
-; GCN-NEXT: v_mov_b32_e32 v46, v6
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB54_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v12
-; GCN-NEXT: v_or_b32_e32 v9, v9, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v13
-; GCN-NEXT: v_or_b32_e32 v11, v11, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v2
-; GCN-NEXT: v_or_b32_e32 v1, v1, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v1
-; GCN-NEXT: v_or_b32_e32 v3, v3, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v1
-; GCN-NEXT: v_or_b32_e32 v4, v5, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v8
-; GCN-NEXT: v_or_b32_e32 v5, v7, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v46
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v31
-; GCN-NEXT: v_or_b32_e32 v6, v10, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v32
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v33
-; GCN-NEXT: v_or_b32_e32 v32, v29, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v34
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v35
-; GCN-NEXT: v_or_b32_e32 v34, v29, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v44, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v46, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v56, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v58, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v61, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v63, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v57, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v62, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v45, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v60, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v41, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v59, v59
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v47, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v43, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v40, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v55
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v36, 0x38000000, v36
-; GCN-NEXT: v_add_f32_e32 v38, 0x38000000, v38
-; GCN-NEXT: v_add_f32_e32 v48, 0x38000000, v48
-; GCN-NEXT: v_add_f32_e32 v49, 0x38000000, v49
-; GCN-NEXT: v_add_f32_e32 v51, 0x38000000, v51
-; GCN-NEXT: v_add_f32_e32 v52, 0x38000000, v52
-; GCN-NEXT: v_add_f32_e32 v54, 0x38000000, v54
-; GCN-NEXT: v_add_f32_e32 v42, 0x38000000, v42
-; GCN-NEXT: v_add_f32_e32 v44, 0x38000000, v44
-; GCN-NEXT: v_add_f32_e32 v46, 0x38000000, v46
-; GCN-NEXT: v_add_f32_e32 v56, 0x38000000, v56
-; GCN-NEXT: v_add_f32_e32 v58, 0x38000000, v58
-; GCN-NEXT: v_add_f32_e32 v61, 0x38000000, v61
-; GCN-NEXT: v_add_f32_e32 v63, 0x38000000, v63
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v57, 0x38000000, v57
-; GCN-NEXT: v_add_f32_e32 v62, 0x38000000, v62
-; GCN-NEXT: v_add_f32_e32 v45, 0x38000000, v45
-; GCN-NEXT: v_add_f32_e32 v60, 0x38000000, v60
-; GCN-NEXT: v_add_f32_e32 v41, 0x38000000, v41
-; GCN-NEXT: v_add_f32_e32 v59, 0x38000000, v59
-; GCN-NEXT: v_add_f32_e32 v53, 0x38000000, v53
-; GCN-NEXT: v_add_f32_e32 v47, 0x38000000, v47
-; GCN-NEXT: v_add_f32_e32 v50, 0x38000000, v50
-; GCN-NEXT: v_add_f32_e32 v43, 0x38000000, v43
-; GCN-NEXT: v_add_f32_e32 v39, 0x38000000, v39
-; GCN-NEXT: v_add_f32_e32 v40, 0x38000000, v40
-; GCN-NEXT: v_add_f32_e32 v37, 0x38000000, v37
-; GCN-NEXT: v_add_f32_e32 v55, 0x38000000, v55
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v44
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v56
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v58
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v61
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v63
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v57
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v62
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v45
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v60
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v59
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v61
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v60
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_or_b32_e32 v30, v30, v29
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v61, v38, v36
-; GCN-NEXT: v_or_b32_e32 v49, v49, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v52, v51
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v42, v42, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v46, v44
-; GCN-NEXT: v_mov_b32_e32 v46, v6
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v58, v56
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v63, v2
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v57, v28
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v45, v45, v62
-; GCN-NEXT: v_or_b32_e32 v41, v41, v60
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v53, v59
-; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v50, v50, v47
-; GCN-NEXT: v_or_b32_e32 v39, v39, v43
-; GCN-NEXT: v_or_b32_e32 v37, v37, v40
-; GCN-NEXT: v_or_b32_e32 v1, v1, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: v_alignbit_b32 v63, v34, v29, 16
-; GCN-NEXT: v_alignbit_b32 v36, v32, v36, 16
-; GCN-NEXT: v_alignbit_b32 v48, v46, v48, 16
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_alignbit_b32 v51, v5, v51, 16
-; GCN-NEXT: v_mov_b32_e32 v7, v4
-; GCN-NEXT: v_alignbit_b32 v54, v4, v54, 16
-; GCN-NEXT: v_alignbit_b32 v29, v3, v44, 16
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(1) expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v29, v1, v56, 16
-; GCN-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v2, v26, v2, 16
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v28, v24, v28, 16
-; GCN-NEXT: v_alignbit_b32 v53, v22, v62, 16
-; GCN-NEXT: v_alignbit_b32 v60, v20, v60, 16
-; GCN-NEXT: v_alignbit_b32 v59, v18, v59, 16
-; GCN-NEXT: v_alignbit_b32 v47, v16, v47, 16
-; GCN-NEXT: v_alignbit_b32 v43, v14, v43, 16
-; GCN-NEXT: v_alignbit_b32 v40, v11, v40, 16
-; GCN-NEXT: v_alignbit_b32 v55, v9, v55, 16
-; GCN-NEXT: .LBB54_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v30, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v63
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v35
-; GCN-NEXT: v_or_b32_e32 v2, v2, v29
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v61
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v36
-; GCN-NEXT: v_or_b32_e32 v56, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v48
-; GCN-NEXT: v_or_b32_e32 v61, v1, v2
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_add_i32_e32 v63, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v51
-; GCN-NEXT: v_or_b32_e32 v57, v1, v2
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v8
-; GCN-NEXT: v_or_b32_e32 v10, v1, v2
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v54
-; GCN-NEXT: v_or_b32_e32 v31, v1, v2
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v7
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v7, v1, v2
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v33, v1, v2
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v5, v1, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v35, v1, v2
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v30
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v1, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v1, v1, v30
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 52, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v30, v28
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 60, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v24, v24, v25
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 64, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v53
-; GCN-NEXT: v_or_b32_e32 v49, v30, v49
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x44, v0
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 0x48, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v60
-; GCN-NEXT: v_or_b32_e32 v45, v30, v54
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 0x4c, v0
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v59
-; GCN-NEXT: v_or_b32_e32 v53, v30, v53
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x54, v0
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x58, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v47
-; GCN-NEXT: v_or_b32_e32 v50, v30, v50
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 0x5c, v0
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 0x60, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v43
-; GCN-NEXT: v_or_b32_e32 v39, v30, v39
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x64, v0
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x68, v0
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v40
-; GCN-NEXT: v_or_b32_e32 v30, v30, v37
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x6c, v0
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v11, v11, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v37, 0xffff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_or_b32_e32 v37, v37, v55
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x74, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v9, v9, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v56, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v61, v63, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v58, v46, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v57, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v34, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v33, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v36, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v48, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v49, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v45, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v53, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v42, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64f16_to_v64i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB54_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 0x200
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
-; VI-NEXT: v_add_f16_sdwa v15, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v33, v15
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v14
-; VI-NEXT: v_add_f16_sdwa v14, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v33, v14
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v13
-; VI-NEXT: v_add_f16_sdwa v13, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v33, v13
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v12
-; VI-NEXT: v_add_f16_sdwa v12, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v33, v12
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v11
-; VI-NEXT: v_add_f16_sdwa v11, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v33, v11
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v10
-; VI-NEXT: v_add_f16_sdwa v10, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v33, v10
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v9
-; VI-NEXT: v_add_f16_sdwa v9, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v33, v9
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v8, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v33, v8
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v33, v7
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v33, v6
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v33, v5
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v33, v4
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v33, v3
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v33, v2
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v33, v1
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v33, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v31
-; VI-NEXT: v_add_f16_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v31, v33, v31
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v30
-; VI-NEXT: v_add_f16_sdwa v30, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v33, v30
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v29
-; VI-NEXT: v_add_f16_sdwa v29, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v33, v29
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v28
-; VI-NEXT: v_add_f16_sdwa v28, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v28, v33, v28
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v27
-; VI-NEXT: v_add_f16_sdwa v27, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v27, v33, v27
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v26
-; VI-NEXT: v_add_f16_sdwa v26, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v26, v33, v26
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v25
-; VI-NEXT: v_add_f16_sdwa v25, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v25, v33, v25
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v24
-; VI-NEXT: v_add_f16_sdwa v24, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v24, v33, v24
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v23
-; VI-NEXT: v_add_f16_sdwa v23, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v23, v33, v23
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v22
-; VI-NEXT: v_add_f16_sdwa v22, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v22, v33, v22
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v21
-; VI-NEXT: v_add_f16_sdwa v21, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v21, v33, v21
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v20
-; VI-NEXT: v_add_f16_sdwa v20, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v20, v33, v20
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v19
-; VI-NEXT: v_add_f16_sdwa v19, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v19, v33, v19
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v18
-; VI-NEXT: v_add_f16_sdwa v18, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v18, v33, v18
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v17
-; VI-NEXT: v_add_f16_sdwa v17, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v17, v33, v17
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB54_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64f16_to_v64i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB54_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB54_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64f16_to_v64i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB54_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB54_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <64 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <64 x half> %a1 to <64 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x half> %a to <64 x i16>
- br label %end
-
-end:
- %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i16> %phi
-}
-
-define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i16_to_v64f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:104
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:40
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt vmcnt(8)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v58
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v52
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v40
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v42
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: .LBB55_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v46
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: v_add_i32_e32 v45, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v44, vcc, 3, v44
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v42, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 3, v55
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v48, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v36, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v46, vcc, 3, v63
-; GCN-NEXT: s_waitcnt vmcnt(8)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31
-; GCN-NEXT: s_waitcnt vmcnt(7)
-; GCN-NEXT: v_add_i32_e32 v61, vcc, 3, v61
-; GCN-NEXT: s_waitcnt vmcnt(6)
-; GCN-NEXT: v_add_i32_e32 v60, vcc, 3, v60
-; GCN-NEXT: s_waitcnt vmcnt(5)
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 3, v59
-; GCN-NEXT: s_waitcnt vmcnt(4)
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 3, v58
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 3, v57
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 3, v56
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v47, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v62, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v63, vcc, 3, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v63
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v5
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v13
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v14
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v18
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v20
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v22
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v26
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v28
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v29
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v30
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v47
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v57
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v58
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v59
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v60
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v61
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v48
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v50
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v52
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v55
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v40
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v41
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v42
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v43
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB55_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v45, v2, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v44, v2, v1
-; GCN-NEXT: v_add_i32_e32 v59, vcc, 8, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v47, v2, v1
-; GCN-NEXT: v_add_i32_e32 v58, vcc, 12, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v46, v2, v1
-; GCN-NEXT: v_add_i32_e32 v57, vcc, 16, v0
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v1
-; GCN-NEXT: v_add_i32_e32 v56, vcc, 20, v0
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v4, v5, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v6, v7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v10, v11, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 40, v0
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v12, v13, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 44, v0
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v18, v19, v18
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v20, v21, v20
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 60, v0
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_or_b32_e32 v22, v23, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 64, v0
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v24, v25, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x44, v0
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v27, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x48, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v28, v29, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x4c, v0
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v30, v31, v30
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x50, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v32, v33, v32
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 0x54, v0
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v34
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v34
-; GCN-NEXT: v_or_b32_e32 v34, v35, v34
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 0x58, v0
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v36
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v36, v37, v36
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 0x5c, v0
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v38
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_or_b32_e32 v38, v39, v38
-; GCN-NEXT: v_add_i32_e32 v39, vcc, 0x60, v0
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v48
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GCN-NEXT: v_or_b32_e32 v48, v49, v48
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 0x64, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v50
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_or_b32_e32 v50, v51, v50
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 0x68, v0
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v52
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_or_b32_e32 v52, v53, v52
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 0x6c, v0
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v54
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_or_b32_e32 v54, v55, v54
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 0x70, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v40
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v40
-; GCN-NEXT: v_or_b32_e32 v40, v41, v40
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 0x74, v0
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v42
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_or_b32_e32 v42, v43, v42
-; GCN-NEXT: v_add_i32_e32 v43, vcc, 0x78, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; GCN-NEXT: buffer_store_dword v45, v59, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v44, v58, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v47, v57, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v46, v56, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v28, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v33, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v32, v35, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v34, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v39, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v48, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v52, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v54, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v40, v43, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v42, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i16_to_v64f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB55_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v32, 3
-; VI-NEXT: v_add_u16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v33
-; VI-NEXT: v_add_u16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v33
-; VI-NEXT: v_add_u16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v33
-; VI-NEXT: v_add_u16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v33
-; VI-NEXT: v_add_u16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v33
-; VI-NEXT: v_add_u16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 3, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v33
-; VI-NEXT: v_add_u16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 3, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v33
-; VI-NEXT: v_add_u16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v33
-; VI-NEXT: v_add_u16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v33
-; VI-NEXT: v_add_u16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v33
-; VI-NEXT: v_add_u16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v33
-; VI-NEXT: v_add_u16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v33
-; VI-NEXT: v_add_u16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v33
-; VI-NEXT: v_add_u16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v33
-; VI-NEXT: v_add_u16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_e32 v1, v1, v33
-; VI-NEXT: v_add_u16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v33
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v31, 3, v31
-; VI-NEXT: v_or_b32_e32 v31, v31, v33
-; VI-NEXT: v_add_u16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v30, 3, v30
-; VI-NEXT: v_or_b32_e32 v30, v30, v33
-; VI-NEXT: v_add_u16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v29, 3, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v33
-; VI-NEXT: v_add_u16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v28, 3, v28
-; VI-NEXT: v_or_b32_e32 v28, v28, v33
-; VI-NEXT: v_add_u16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v27, 3, v27
-; VI-NEXT: v_or_b32_e32 v27, v27, v33
-; VI-NEXT: v_add_u16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v26, 3, v26
-; VI-NEXT: v_or_b32_e32 v26, v26, v33
-; VI-NEXT: v_add_u16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v25, 3, v25
-; VI-NEXT: v_or_b32_e32 v25, v25, v33
-; VI-NEXT: v_add_u16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v24, 3, v24
-; VI-NEXT: v_or_b32_e32 v24, v24, v33
-; VI-NEXT: v_add_u16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v23, 3, v23
-; VI-NEXT: v_or_b32_e32 v23, v23, v33
-; VI-NEXT: v_add_u16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v22, 3, v22
-; VI-NEXT: v_or_b32_e32 v22, v22, v33
-; VI-NEXT: v_add_u16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v21, 3, v21
-; VI-NEXT: v_or_b32_e32 v21, v21, v33
-; VI-NEXT: v_add_u16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v20, 3, v20
-; VI-NEXT: v_or_b32_e32 v20, v20, v33
-; VI-NEXT: v_add_u16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v19, 3, v19
-; VI-NEXT: v_or_b32_e32 v19, v19, v33
-; VI-NEXT: v_add_u16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v18, 3, v18
-; VI-NEXT: v_or_b32_e32 v18, v18, v33
-; VI-NEXT: v_add_u16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_add_u16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v16
-; VI-NEXT: v_or_b32_e32 v17, v17, v33
-; VI-NEXT: v_or_b32_e32 v16, v16, v32
-; VI-NEXT: .LBB55_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i16_to_v64f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB55_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB55_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i16_to_v64f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v32
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB55_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB55_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i16> %a, splat (i16 3)
- %a2 = bitcast <64 x i16> %a1 to <64 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i16> %a to <64 x half>
- br label %end
-
-end:
- %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x half> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
deleted file mode 100644
index 03c6a36ac9861..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
+++ /dev/null
@@ -1,11387 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <4 x float> @bitcast_v4i32_to_v4f32(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <4 x i32> @bitcast_v4f32_to_v4i32(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <2 x i64> @bitcast_v4i32_to_v2i64(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <4 x i32> @bitcast_v2i64_to_v4i32(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <2 x double> @bitcast_v4i32_to_v2f64(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <4 x i32> @bitcast_v2f64_to_v4i32(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <8 x i16> @bitcast_v4i32_to_v8i16(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v8
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <4 x i32> @bitcast_v8i16_to_v4i32(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_4
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB7_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v0, v8
-; GCN-NEXT: v_or_b32_e32 v1, v1, v11
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: .LBB7_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v8, v0
-; GCN-NEXT: v_or_b32_e32 v1, v11, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x30000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 3
-; VI-NEXT: v_add_u16_e32 v4, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: v_add_u16_e32 v4, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v4, v2
-; VI-NEXT: v_add_u16_e32 v4, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v4, v1
-; VI-NEXT: v_add_u16_e32 v4, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <8 x half> @bitcast_v4i32_to_v8f16(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v9, v3
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v11, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB8_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB8_4
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB8_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: .LBB8_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <4 x i32> @bitcast_v8f16_to_v4i32(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB9_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB9_4
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB9_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v10, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: .LBB9_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v6
-; GCN-NEXT: v_or_b32_e32 v3, v4, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 0x200
-; VI-NEXT: v_add_f16_sdwa v5, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v5
-; VI-NEXT: v_add_f16_sdwa v5, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v5
-; VI-NEXT: v_add_f16_sdwa v5, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v4, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v5
-; VI-NEXT: v_or_b32_e32 v0, v0, v4
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <8 x bfloat> @bitcast_v4i32_to_v8bf16(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v11, v3
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_4
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB10_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v10
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: .LBB10_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v11
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <4 x i32> @bitcast_v8bf16_to_v4i32(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_4
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB11_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v0, v11, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; GCN-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: .LBB11_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v2, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v3, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v3
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v2
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v1
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v1
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v2, v4, v2, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v1, v4, v1, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s7
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_add3_u32 v9, v9, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX11-NEXT: v_add3_u32 v7, v13, v2, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v7, v8 :: v_dual_add_f32 v7, 0x40c00000, v10
-; GFX11-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v8, v9, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v6
-; GFX11-NEXT: v_bfe_u32 v11, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
-; GFX11-NEXT: v_add3_u32 v11, v11, v7, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-NEXT: v_add3_u32 v9, v10, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_cndmask_b32 v1, v9, v10
-; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_perm_b32 v1, v6, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v8, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v7, v0, 0x7060302
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <16 x i8> @bitcast_v4i32_to_v16i8(<4 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i32_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v16, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB12_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB12_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i32_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB12_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB12_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i32_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB12_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
-; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB12_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i32_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB12_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
-; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB12_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i32> %a, splat (i32 3)
- %a2 = bitcast <4 x i32> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i32> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v4i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_4
-; GCN-NEXT: .LBB13_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB13_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v0, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: .LBB13_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v0, v19, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v20, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v21, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x300, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v4i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB13_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB13_4
-; VI-NEXT: .LBB13_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB13_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_2
-; VI-NEXT: .LBB13_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v4i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB13_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB13_4
-; GFX9-NEXT: .LBB13_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB13_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_2
-; GFX9-NEXT: .LBB13_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v4i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_4
-; GFX11-NEXT: .LBB13_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB13_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB13_2
-; GFX11-NEXT: .LBB13_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <4 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <4 x i32>
- br label %end
-
-end:
- %phi = phi <4 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i32> %phi
-}
-
-define <2 x i64> @bitcast_v4f32_to_v2i64(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <4 x float> @bitcast_v2i64_to_v4f32(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <2 x double> @bitcast_v4f32_to_v2f64(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <4 x float> @bitcast_v2f64_to_v4f32(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <8 x i16> @bitcast_v4f32_to_v8i16(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v8
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <4 x float> @bitcast_v8i16_to_v4f32(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_4
-; GCN-NEXT: .LBB19_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB19_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v0, v8
-; GCN-NEXT: v_or_b32_e32 v1, v1, v11
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: .LBB19_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v8, v0
-; GCN-NEXT: v_or_b32_e32 v1, v11, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x30000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 3
-; VI-NEXT: v_add_u16_e32 v4, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: v_add_u16_e32 v4, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v4, v2
-; VI-NEXT: v_add_u16_e32 v4, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v4, v1
-; VI-NEXT: v_add_u16_e32 v4, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB19_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <8 x half> @bitcast_v4f32_to_v8f16(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v9, v3
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v11, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_4
-; GCN-NEXT: .LBB20_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB20_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: .LBB20_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <4 x float> @bitcast_v8f16_to_v4f32(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_4
-; GCN-NEXT: .LBB21_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB21_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v10, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: .LBB21_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v6
-; GCN-NEXT: v_or_b32_e32 v3, v4, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 0x200
-; VI-NEXT: v_add_f16_sdwa v5, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v5
-; VI-NEXT: v_add_f16_sdwa v5, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v5
-; VI-NEXT: v_add_f16_sdwa v5, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v4, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v5
-; VI-NEXT: v_or_b32_e32 v0, v0, v4
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <8 x bfloat> @bitcast_v4f32_to_v8bf16(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v11, v3
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_4
-; GCN-NEXT: .LBB22_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB22_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v10
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: .LBB22_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v11
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <4 x float> @bitcast_v8bf16_to_v4f32(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_4
-; GCN-NEXT: .LBB23_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB23_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v0, v11, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; GCN-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: .LBB23_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v2, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB23_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v3, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v3
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v2
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v1
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v1
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB23_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB23_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v2, v4, v2, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v1, v4, v1, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s7
-; GFX9-NEXT: .LBB23_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB23_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_add3_u32 v9, v9, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX11-NEXT: v_add3_u32 v7, v13, v2, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v7, v8 :: v_dual_add_f32 v7, 0x40c00000, v10
-; GFX11-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v8, v9, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v6
-; GFX11-NEXT: v_bfe_u32 v11, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
-; GFX11-NEXT: v_add3_u32 v11, v11, v7, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-NEXT: v_add3_u32 v9, v10, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_cndmask_b32 v1, v9, v10
-; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_perm_b32 v1, v6, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v8, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v7, v0, 0x7060302
-; GFX11-NEXT: .LBB23_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <16 x i8> @bitcast_v4f32_to_v16i8(<4 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f32_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v16, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB24_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB24_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f32_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB24_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
-; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
-; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB24_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f32_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB24_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
-; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB24_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f32_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB24_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB24_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <4 x float> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x float> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v4f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_4
-; GCN-NEXT: .LBB25_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB25_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v0, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: .LBB25_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v0, v19, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v20, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v21, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x300, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v4f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB25_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB25_4
-; VI-NEXT: .LBB25_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB25_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: .LBB25_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v4f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB25_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB25_4
-; GFX9-NEXT: .LBB25_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB25_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: .LBB25_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v4f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_4
-; GFX11-NEXT: .LBB25_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB25_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: .LBB25_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <4 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <4 x float>
- br label %end
-
-end:
- %phi = phi <4 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x float> %phi
-}
-
-define <2 x double> @bitcast_v2i64_to_v2f64(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: .LBB26_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <2 x i64> @bitcast_v2f64_to_v2i64(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: .LBB27_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: .LBB27_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: .LBB27_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: .LBB27_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <8 x i16> @bitcast_v2i64_to_v8i16(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_alignbit_b32 v5, v6, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v8
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <2 x i64> @bitcast_v8i16_to_v2i64(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_4
-; GCN-NEXT: .LBB29_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB29_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v0, v8
-; GCN-NEXT: v_or_b32_e32 v1, v1, v11
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: .LBB29_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v8, v0
-; GCN-NEXT: v_or_b32_e32 v1, v11, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x30000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 3
-; VI-NEXT: v_add_u16_e32 v4, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: v_add_u16_e32 v4, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v4, v2
-; VI-NEXT: v_add_u16_e32 v4, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v4, v1
-; VI-NEXT: v_add_u16_e32 v4, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB29_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <8 x half> @bitcast_v2i64_to_v8f16(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v3
-; GCN-NEXT: v_mov_b32_e32 v9, v2
-; GCN-NEXT: v_mov_b32_e32 v11, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_4
-; GCN-NEXT: .LBB30_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB30_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_2
-; GCN-NEXT: .LBB30_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v10, vcc
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <2 x i64> @bitcast_v8f16_to_v2i64(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_4
-; GCN-NEXT: .LBB31_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB31_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v10, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_2
-; GCN-NEXT: .LBB31_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v6
-; GCN-NEXT: v_or_b32_e32 v3, v4, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB31_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 0x200
-; VI-NEXT: v_add_f16_sdwa v5, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v5
-; VI-NEXT: v_add_f16_sdwa v5, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v5
-; VI-NEXT: v_add_f16_sdwa v5, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v4, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v5
-; VI-NEXT: v_or_b32_e32 v0, v0, v4
-; VI-NEXT: .LBB31_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <8 x bfloat> @bitcast_v2i64_to_v8bf16(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v11, v3
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_4
-; GCN-NEXT: .LBB32_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB32_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v10
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v8
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_2
-; GCN-NEXT: .LBB32_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v11, vcc
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <2 x i64> @bitcast_v8bf16_to_v2i64(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_4
-; GCN-NEXT: .LBB33_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB33_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v0, v11, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; GCN-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_2
-; GCN-NEXT: .LBB33_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v2, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB33_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v3, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v3
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v2
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v1
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v1
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB33_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB33_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v2, v4, v2, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v1, v4, v1, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s7
-; GFX9-NEXT: .LBB33_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB33_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_add3_u32 v9, v9, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX11-NEXT: v_add3_u32 v7, v13, v2, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v7, v8 :: v_dual_add_f32 v7, 0x40c00000, v10
-; GFX11-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v8, v9, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v6
-; GFX11-NEXT: v_bfe_u32 v11, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
-; GFX11-NEXT: v_add3_u32 v11, v11, v7, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-NEXT: v_add3_u32 v9, v10, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_cndmask_b32 v1, v9, v10
-; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_perm_b32 v1, v6, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v8, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v7, v0, 0x7060302
-; GFX11-NEXT: .LBB33_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <16 x i8> @bitcast_v2i64_to_v16i8(<2 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i64_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v16, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB34_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v3, v16, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v16, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v16, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v16
-; GCN-NEXT: .LBB34_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i64_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB34_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
-; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB34_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i64_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB34_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
-; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
-; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB34_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i64_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB34_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB34_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i64> %a, splat (i64 3)
- %a2 = bitcast <2 x i64> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i64> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v2i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_4
-; GCN-NEXT: .LBB35_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB35_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v0, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_2
-; GCN-NEXT: .LBB35_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v0, v19, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v20, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v21, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x300, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v2i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_4
-; VI-NEXT: .LBB35_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB35_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_2
-; VI-NEXT: .LBB35_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v2i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_4
-; GFX9-NEXT: .LBB35_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB35_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_2
-; GFX9-NEXT: .LBB35_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v2i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_4
-; GFX11-NEXT: .LBB35_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB35_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB35_2
-; GFX11-NEXT: .LBB35_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <2 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <2 x i64>
- br label %end
-
-end:
- %phi = phi <2 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i64> %phi
-}
-
-define <8 x i16> @bitcast_v2f64_to_v8i16(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v9, v3
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v11, v1
-; GCN-NEXT: v_mov_b32_e32 v10, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v11, v10, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: .LBB36_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v1, v11, v10, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: .LBB36_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v10
-; GCN-NEXT: v_mov_b32_e32 v2, v11
-; GCN-NEXT: v_mov_b32_e32 v4, v8
-; GCN-NEXT: v_mov_b32_e32 v6, v9
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB36_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB36_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB36_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <2 x double> @bitcast_v8i16_to_v2f64(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v10, v2
-; GCN-NEXT: v_mov_b32_e32 v9, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_4
-; GCN-NEXT: .LBB37_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB37_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v0, v8
-; GCN-NEXT: v_or_b32_e32 v1, v1, v11
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_2
-; GCN-NEXT: .LBB37_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v8, v0
-; GCN-NEXT: v_or_b32_e32 v1, v11, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x30000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB37_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 3
-; VI-NEXT: v_add_u16_e32 v4, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: v_add_u16_e32 v4, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v4, v2
-; VI-NEXT: v_add_u16_e32 v4, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v4, v1
-; VI-NEXT: v_add_u16_e32 v4, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB37_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <8 x half> @bitcast_v2f64_to_v8f16(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: .LBB38_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: .LBB38_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v9
-; GCN-NEXT: v_mov_b32_e32 v1, v11
-; GCN-NEXT: v_mov_b32_e32 v2, v8
-; GCN-NEXT: v_mov_b32_e32 v3, v10
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB38_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB38_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB38_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB38_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB38_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB38_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <2 x double> @bitcast_v8f16_to_v2f64(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_4
-; GCN-NEXT: .LBB39_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB39_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v10, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_2
-; GCN-NEXT: .LBB39_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v6
-; GCN-NEXT: v_or_b32_e32 v3, v4, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB39_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 0x200
-; VI-NEXT: v_add_f16_sdwa v5, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v5
-; VI-NEXT: v_add_f16_sdwa v5, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v5
-; VI-NEXT: v_add_f16_sdwa v5, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v4, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v5
-; VI-NEXT: v_or_b32_e32 v0, v0, v4
-; VI-NEXT: .LBB39_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <8 x bfloat> @bitcast_v2f64_to_v8bf16(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: .LBB40_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v0
-; GCN-NEXT: .LBB40_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v11
-; GCN-NEXT: v_mov_b32_e32 v1, v10
-; GCN-NEXT: v_mov_b32_e32 v2, v9
-; GCN-NEXT: v_mov_b32_e32 v3, v8
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB40_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB40_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB40_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB40_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB40_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB40_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <2 x double> @bitcast_v8bf16_to_v2f64(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v6
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_4
-; GCN-NEXT: .LBB41_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB41_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v0, v11, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; GCN-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_2
-; GCN-NEXT: .LBB41_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v2, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB41_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v3, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v3
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v2
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v1
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v1
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v4, 16
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB41_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB41_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v2, v4, v2, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v1, v4, v1, s7
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s7
-; GFX9-NEXT: .LBB41_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB41_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_add3_u32 v9, v9, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX11-NEXT: v_add3_u32 v7, v13, v2, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v7, v8 :: v_dual_add_f32 v7, 0x40c00000, v10
-; GFX11-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v8, v9, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v6
-; GFX11-NEXT: v_bfe_u32 v11, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
-; GFX11-NEXT: v_add3_u32 v11, v11, v7, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-NEXT: v_add3_u32 v9, v10, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_cndmask_b32 v1, v9, v10
-; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_perm_b32 v1, v6, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v8, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v7, v0, 0x7060302
-; GFX11-NEXT: .LBB41_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <16 x i8> @bitcast_v2f64_to_v16i8(<2 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f64_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v3
-; GCN-NEXT: v_mov_b32_e32 v16, v2
-; GCN-NEXT: v_mov_b32_e32 v19, v1
-; GCN-NEXT: v_mov_b32_e32 v18, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v17, v16, 24
-; GCN-NEXT: v_alignbit_b32 v10, v17, v16, 16
-; GCN-NEXT: v_alignbit_b32 v9, v17, v16, 8
-; GCN-NEXT: v_alignbit_b32 v3, v19, v18, 24
-; GCN-NEXT: v_alignbit_b32 v2, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v1, v19, v18, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GCN-NEXT: .LBB42_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GCN-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GCN-NEXT: v_alignbit_b32 v11, v17, v16, 24
-; GCN-NEXT: v_alignbit_b32 v10, v17, v16, 16
-; GCN-NEXT: v_alignbit_b32 v9, v17, v16, 8
-; GCN-NEXT: v_alignbit_b32 v3, v19, v18, 24
-; GCN-NEXT: v_alignbit_b32 v2, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v1, v19, v18, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GCN-NEXT: .LBB42_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v18
-; GCN-NEXT: v_mov_b32_e32 v4, v19
-; GCN-NEXT: v_mov_b32_e32 v8, v16
-; GCN-NEXT: v_mov_b32_e32 v12, v17
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f64_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB42_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB42_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f64_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB42_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB42_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f64_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB42_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB42_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <2 x double> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x double> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v2f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB43_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB43_4
-; GCN-NEXT: .LBB43_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB43_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v0, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_2
-; GCN-NEXT: .LBB43_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_or_b32_e32 v0, v19, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v20, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v21, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v16, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v7, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x300, v8
-; GCN-NEXT: v_or_b32_e32 v7, v9, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v2f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB43_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB43_4
-; VI-NEXT: .LBB43_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB43_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB43_2
-; VI-NEXT: .LBB43_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v2f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB43_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB43_4
-; GFX9-NEXT: .LBB43_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB43_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB43_2
-; GFX9-NEXT: .LBB43_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v2f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_4
-; GFX11-NEXT: .LBB43_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB43_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB43_2
-; GFX11-NEXT: .LBB43_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <2 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <2 x double>
- br label %end
-
-end:
- %phi = phi <2 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x double> %phi
-}
-
-define <8 x half> @bitcast_v8i16_to_v8f16(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v16, v7
-; GCN-NEXT: v_mov_b32_e32 v9, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v11, v4
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v13, v2
-; GCN-NEXT: v_mov_b32_e32 v14, v1
-; GCN-NEXT: v_mov_b32_e32 v15, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB44_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB44_4
-; GCN-NEXT: .LBB44_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB44_3: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_2
-; GCN-NEXT: .LBB44_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB44_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 3
-; VI-NEXT: v_add_u16_sdwa v5, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v6, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v7, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v4, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_or_b32_e32 v2, v2, v7
-; VI-NEXT: v_or_b32_e32 v1, v1, v6
-; VI-NEXT: v_or_b32_e32 v0, v0, v5
-; VI-NEXT: .LBB44_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <8 x i16> @bitcast_v8f16_to_v8i16(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB45_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v8
-; GCN-NEXT: v_or_b32_e32 v2, v2, v9
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: .LBB45_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB45_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 0x200
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_or_b32_e32 v2, v7, v2
-; VI-NEXT: v_or_b32_e32 v1, v6, v1
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB45_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <8 x bfloat> @bitcast_v8i16_to_v8bf16(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v11, v6
-; GCN-NEXT: v_mov_b32_e32 v12, v4
-; GCN-NEXT: v_mov_b32_e32 v9, v2
-; GCN-NEXT: v_mov_b32_e32 v10, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_4
-; GCN-NEXT: .LBB46_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB46_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v11
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB46_2
-; GCN-NEXT: .LBB46_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v11
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v0, v7, v0
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_or_b32_e32 v1, v1, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB46_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v4, 3
-; VI-NEXT: v_add_u16_sdwa v5, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v6, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v7, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v4, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_or_b32_e32 v2, v2, v7
-; VI-NEXT: v_or_b32_e32 v1, v1, v6
-; VI-NEXT: v_or_b32_e32 v0, v0, v5
-; VI-NEXT: .LBB46_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <8 x i16> @bitcast_v8bf16_to_v8i16(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v7
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB47_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB47_4
-; GCN-NEXT: .LBB47_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB47_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v10
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_2
-; GCN-NEXT: .LBB47_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v14
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v4
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_alignbit_b32 v0, v9, v0, 16
-; GCN-NEXT: v_alignbit_b32 v4, v10, v2, 16
-; GCN-NEXT: v_alignbit_b32 v6, v7, v8, 16
-; GCN-NEXT: v_alignbit_b32 v2, v3, v5, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v11, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB47_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v1
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_bfe_u32 v6, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v5
-; VI-NEXT: v_add_u32_e32 v6, vcc, s6, v6
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc
-; VI-NEXT: v_bfe_u32 v6, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v1
-; VI-NEXT: v_add_u32_e32 v6, vcc, s6, v6
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v6, 16, v2
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_bfe_u32 v7, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v7, vcc, v7, v6
-; VI-NEXT: v_add_u32_e32 v7, vcc, s6, v7
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v8, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
-; VI-NEXT: v_bfe_u32 v7, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; VI-NEXT: v_add_u32_e32 v7, vcc, s6, v7
-; VI-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v3
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_bfe_u32 v8, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v7
-; VI-NEXT: v_add_u32_e32 v8, vcc, s6, v8
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v7, v8, v9, vcc
-; VI-NEXT: v_bfe_u32 v8, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v3
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v7, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v6, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v5, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB47_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB47_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_bfe_u32 v6, v5, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v6, v6, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc
-; GFX9-NEXT: v_bfe_u32 v6, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v6, v6, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v7, v7, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
-; GFX9-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v7, v7, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_bfe_u32 v8, v7, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v8, v8, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v8, v9, vcc
-; GFX9-NEXT: v_bfe_u32 v8, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v8, v8, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v7, v3, s6
-; GFX9-NEXT: v_perm_b32 v2, v6, v2, s6
-; GFX9-NEXT: v_perm_b32 v1, v5, v1, s6
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s6
-; GFX9-NEXT: .LBB47_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB47_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_and_b32 v5, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v13, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_add3_u32 v7, v13, v1, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_bfe_u32 v11, v6, 16, 1
-; GFX11-NEXT: v_add3_u32 v9, v9, v0, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v7, v8 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v8, v11, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_bfe_u32 v12, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v13, v7, 16, 1
-; GFX11-NEXT: v_bfe_u32 v14, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v12, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v11, v13, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v13, v14, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
-; GFX11-NEXT: v_perm_b32 v1, v5, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v13, v14, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_perm_b32 v3, v7, v3, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v8, v15, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v4, v0, 0x7060302
-; GFX11-NEXT: .LBB47_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <16 x i8> @bitcast_v8i16_to_v16i8(<8 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i16_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v7
-; GCN-NEXT: v_mov_b32_e32 v16, v6
-; GCN-NEXT: v_mov_b32_e32 v17, v4
-; GCN-NEXT: v_mov_b32_e32 v21, v3
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v19, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v20
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB48_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB48_4
-; GCN-NEXT: .LBB48_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB48_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v20
-; GCN-NEXT: v_bfe_u32 v7, v21, 8, 8
-; GCN-NEXT: v_or_b32_e32 v0, v0, v24
-; GCN-NEXT: v_or_b32_e32 v4, v1, v25
-; GCN-NEXT: v_or_b32_e32 v8, v2, v22
-; GCN-NEXT: v_or_b32_e32 v12, v3, v23
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_bfe_u32 v15, v20, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_2
-; GCN-NEXT: .LBB48_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v18
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v2, v24, v2
-; GCN-NEXT: v_or_b32_e32 v3, v25, v3
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v3
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i16_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v19, v3
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v18
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v20, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_mov_b32_e32 v16, v0
-; VI-NEXT: v_mov_b32_e32 v17, v1
-; VI-NEXT: v_mov_b32_e32 v8, v18
-; VI-NEXT: v_mov_b32_e32 v21, v19
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: .LBB48_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v3, 3
-; VI-NEXT: v_add_u16_sdwa v6, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v2, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v14, v19, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v10, v18, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v1
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; VI-NEXT: v_add_u16_e32 v16, 3, v0
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; VI-NEXT: v_add_u16_e32 v21, 3, v19
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v14
-; VI-NEXT: v_add_u16_e32 v8, 3, v18
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v10
-; VI-NEXT: v_or_b32_e32 v1, v17, v1
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: v_or_b32_e32 v19, v21, v4
-; VI-NEXT: v_or_b32_e32 v18, v8, v3
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v18
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v20, 8, v0
-; VI-NEXT: v_bfe_u32 v15, v14, 8, 8
-; VI-NEXT: v_bfe_u32 v7, v6, 8, 8
-; VI-NEXT: .LBB48_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v16
-; VI-NEXT: v_mov_b32_e32 v1, v20
-; VI-NEXT: v_mov_b32_e32 v4, v17
-; VI-NEXT: v_mov_b32_e32 v12, v21
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i16_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB48_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB48_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i16_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB48_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB48_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i16> %a, splat (i16 3)
- %a2 = bitcast <8 x i16> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i16> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v8i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v9
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GCN-NEXT: v_or_b32_e32 v1, v1, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v0, v0, v22
-; GCN-NEXT: v_or_b32_e32 v7, v7, v23
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v3, v16, v3
-; GCN-NEXT: v_or_b32_e32 v2, v17, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v8, v18, v5
-; GCN-NEXT: v_or_b32_e32 v5, v19, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v11, v1, v3
-; GCN-NEXT: v_or_b32_e32 v15, v4, v8
-; GCN-NEXT: v_or_b32_e32 v9, v0, v2
-; GCN-NEXT: v_or_b32_e32 v13, v6, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v1, v11, v2, 16
-; GCN-NEXT: v_alignbit_b32 v5, v15, v5, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: .LBB49_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v8
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v10
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v5, v21, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v4, v20, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
-; GCN-NEXT: v_or_b32_e32 v3, v19, v3
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_or_b32_e32 v7, v18, v7
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_or_b32_e32 v2, v17, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v6, v16, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v1, v3, v1
-; GCN-NEXT: v_or_b32_e32 v3, v7, v5
-; GCN-NEXT: v_or_b32_e32 v0, v2, v0
-; GCN-NEXT: v_or_b32_e32 v2, v6, v4
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 0x3000000, v2
-; GCN-NEXT: v_alignbit_b32 v1, v11, v9, 16
-; GCN-NEXT: v_alignbit_b32 v5, v15, v13, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v15
-; GCN-NEXT: .LBB49_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v9
-; GCN-NEXT: v_mov_b32_e32 v2, v11
-; GCN-NEXT: v_mov_b32_e32 v4, v13
-; GCN-NEXT: v_mov_b32_e32 v6, v15
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v8i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB49_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB49_4
-; VI-NEXT: .LBB49_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB49_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_2
-; VI-NEXT: .LBB49_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v8i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB49_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB49_4
-; GFX9-NEXT: .LBB49_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB49_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_2
-; GFX9-NEXT: .LBB49_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v8i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_4
-; GFX11-NEXT: .LBB49_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB49_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB49_2
-; GFX11-NEXT: .LBB49_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <8 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <8 x i16>
- br label %end
-
-end:
- %phi = phi <8 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i16> %phi
-}
-
-define <8 x bfloat> @bitcast_v8f16_to_v8bf16(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v7
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB50_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB50_4
-; GCN-NEXT: .LBB50_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB50_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_2
-; GCN-NEXT: .LBB50_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v8
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v11
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB50_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 0x200
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_or_b32_e32 v2, v7, v2
-; VI-NEXT: v_or_b32_e32 v1, v6, v1
-; VI-NEXT: v_or_b32_e32 v0, v4, v0
-; VI-NEXT: .LBB50_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
-
-define <8 x half> @bitcast_v8bf16_to_v8f16(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v7
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB51_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB51_4
-; GCN-NEXT: .LBB51_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB51_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_2
-; GCN-NEXT: .LBB51_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v15
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v14
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v12
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v8
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB51_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v0
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_bfe_u32 v5, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v4
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; VI-NEXT: v_bfe_u32 v5, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v5, vcc, v5, v0
-; VI-NEXT: v_add_u32_e32 v5, vcc, s6, v5
-; VI-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v1
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_bfe_u32 v6, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v5
-; VI-NEXT: v_add_u32_e32 v6, vcc, s6, v6
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc
-; VI-NEXT: v_bfe_u32 v6, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v1
-; VI-NEXT: v_add_u32_e32 v6, vcc, s6, v6
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v6, 16, v2
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_bfe_u32 v7, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v7, vcc, v7, v6
-; VI-NEXT: v_add_u32_e32 v7, vcc, s6, v7
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v8, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
-; VI-NEXT: v_bfe_u32 v7, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; VI-NEXT: v_add_u32_e32 v7, vcc, s6, v7
-; VI-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v3
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_bfe_u32 v8, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v7
-; VI-NEXT: v_add_u32_e32 v8, vcc, s6, v8
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v7, v8, v9, vcc
-; VI-NEXT: v_bfe_u32 v8, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v3
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v7, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v6, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v5, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v4, 16
-; VI-NEXT: .LBB51_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB51_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_bfe_u32 v6, v5, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add3_u32 v6, v6, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc
-; GFX9-NEXT: v_bfe_u32 v6, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v6, v6, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT: v_add3_u32 v7, v7, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
-; GFX9-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v7, v7, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_bfe_u32 v8, v7, 16, 1
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX9-NEXT: v_add3_u32 v8, v8, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v8, v9, vcc
-; GFX9-NEXT: v_bfe_u32 v8, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v8, v8, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v3, v7, v3, s6
-; GFX9-NEXT: v_perm_b32 v2, v6, v2, s6
-; GFX9-NEXT: v_perm_b32 v1, v5, v1, s6
-; GFX9-NEXT: v_perm_b32 v0, v4, v0, s6
-; GFX9-NEXT: .LBB51_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB51_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v7, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_and_b32 v5, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v0, 16, v0
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v13, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v5, v11, v12 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_add3_u32 v7, v13, v1, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_bfe_u32 v11, v6, 16, 1
-; GFX11-NEXT: v_add3_u32 v9, v9, v0, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v7, v8 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v8, v11, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_bfe_u32 v12, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v13, v7, 16, 1
-; GFX11-NEXT: v_bfe_u32 v14, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v12, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v11, v13, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v13, v14, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
-; GFX11-NEXT: v_perm_b32 v1, v5, v1, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v13, v14, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_perm_b32 v3, v7, v3, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v8, v15, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v4, v0, 0x7060302
-; GFX11-NEXT: .LBB51_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <16 x i8> @bitcast_v8f16_to_v16i8(<8 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f16_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v9, v6
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v9
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB52_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB52_4
-; GCN-NEXT: .LBB52_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB52_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v14
-; GCN-NEXT: v_bfe_u32 v7, v6, 8, 8
-; GCN-NEXT: v_or_b32_e32 v0, v17, v0
-; GCN-NEXT: v_or_b32_e32 v4, v16, v1
-; GCN-NEXT: v_or_b32_e32 v8, v19, v2
-; GCN-NEXT: v_or_b32_e32 v12, v18, v3
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_bfe_u32 v15, v14, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB52_2
-; GCN-NEXT: .LBB52_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v6
-; GCN-NEXT: v_bfe_u32 v7, v6, 8, 8
-; GCN-NEXT: v_or_b32_e32 v8, v1, v0
-; GCN-NEXT: v_or_b32_e32 v12, v2, v9
-; GCN-NEXT: v_or_b32_e32 v0, v4, v3
-; GCN-NEXT: v_or_b32_e32 v4, v5, v10
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_bfe_u32 v15, v14, 8, 8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f16_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB52_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB52_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB52_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v3, 0x200
-; VI-NEXT: v_add_f16_sdwa v6, v19, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
-; VI-NEXT: v_add_f16_sdwa v2, v18, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v14, v17, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v10, v16, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v19, v0
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v14
-; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v10
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
-; VI-NEXT: v_or_b32_e32 v0, v18, v0
-; VI-NEXT: v_or_b32_e32 v8, v17, v4
-; VI-NEXT: v_or_b32_e32 v7, v16, v3
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: v_bfe_u32 v15, v14, 8, 8
-; VI-NEXT: v_bfe_u32 v7, v6, 8, 8
-; VI-NEXT: .LBB52_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f16_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB52_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB52_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB52_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB52_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f16_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB52_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB52_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB52_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB52_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <8 x half> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x half> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v8f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 8, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v0, v16
-; GCN-NEXT: v_or_b32_e32 v1, v1, v18
-; GCN-NEXT: v_or_b32_e32 v2, v2, v19
-; GCN-NEXT: v_or_b32_e32 v3, v3, v20
-; GCN-NEXT: v_or_b32_e32 v4, v4, v21
-; GCN-NEXT: v_or_b32_e32 v5, v5, v22
-; GCN-NEXT: v_or_b32_e32 v6, v6, v23
-; GCN-NEXT: v_or_b32_e32 v7, v7, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: .LBB53_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v14
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v17
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v15, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v3, v22, v3
-; GCN-NEXT: v_or_b32_e32 v5, v21, v5
-; GCN-NEXT: v_or_b32_e32 v6, v20, v6
-; GCN-NEXT: v_or_b32_e32 v4, v19, v4
-; GCN-NEXT: v_or_b32_e32 v2, v18, v2
-; GCN-NEXT: v_or_b32_e32 v7, v16, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x300, v0
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v8
-; GCN-NEXT: .LBB53_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v2, v11
-; GCN-NEXT: v_mov_b32_e32 v4, v9
-; GCN-NEXT: v_mov_b32_e32 v6, v13
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v8f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB53_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB53_4
-; VI-NEXT: .LBB53_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB53_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB53_2
-; VI-NEXT: .LBB53_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v8f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB53_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB53_4
-; GFX9-NEXT: .LBB53_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB53_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB53_2
-; GFX9-NEXT: .LBB53_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v8f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB53_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB53_4
-; GFX11-NEXT: .LBB53_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB53_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB53_2
-; GFX11-NEXT: .LBB53_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <8 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <8 x half>
- br label %end
-
-end:
- %phi = phi <8 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x half> %phi
-}
-
-define <16 x i8> @bitcast_v8bf16_to_v16i8(<8 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8bf16_to_v16i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v6
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB54_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB54_4
-; GCN-NEXT: .LBB54_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB54_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v20
-; GCN-NEXT: v_alignbit_b32 v0, v0, v18, 16
-; GCN-NEXT: v_alignbit_b32 v4, v6, v16, 16
-; GCN-NEXT: v_alignbit_b32 v8, v1, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v14, v19, 16
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB54_2
-; GCN-NEXT: .LBB54_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v17
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v7
-; GCN-NEXT: v_alignbit_b32 v8, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v12, v14, v2, 16
-; GCN-NEXT: v_alignbit_b32 v0, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v4, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8bf16_to_v16i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v17, v3
-; VI-NEXT: v_mov_b32_e32 v16, v2
-; VI-NEXT: v_mov_b32_e32 v19, v1
-; VI-NEXT: v_mov_b32_e32 v18, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB54_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB54_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB54_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v19
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v19
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, s6, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v19, v1, v0, 16
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v18
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, s6, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v18
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, s6, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v18, v1, v0, 16
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v17
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, s6, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v17
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, s6, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v17, v1, v0, 16
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, s6, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v16
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, s6, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v16, v1, v0, 16
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; VI-NEXT: .LBB54_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v18
-; VI-NEXT: v_mov_b32_e32 v4, v19
-; VI-NEXT: v_mov_b32_e32 v8, v16
-; VI-NEXT: v_mov_b32_e32 v12, v17
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8bf16_to_v16i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v17, v3
-; GFX9-NEXT: v_mov_b32_e32 v16, v2
-; GFX9-NEXT: v_mov_b32_e32 v19, v1
-; GFX9-NEXT: v_mov_b32_e32 v18, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB54_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX9-NEXT: .LBB54_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB54_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v19
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v19
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v1, v2, vcc
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v18
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v1, v3, vcc
-; GFX9-NEXT: v_add3_u32 v4, v4, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v17
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_bfe_u32 v6, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v6, v6, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v17
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v7, v7, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v7, v8, vcc
-; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v7, v7, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v10, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v16
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_bfe_u32 v10, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v10, v11, vcc
-; GFX9-NEXT: v_perm_b32 v1, v2, v3, s7
-; GFX9-NEXT: v_perm_b32 v0, v0, v4, s7
-; GFX9-NEXT: v_perm_b32 v8, v5, v9, s7
-; GFX9-NEXT: v_perm_b32 v7, v6, v10, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v4
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: .LBB54_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v0, v18
-; GFX9-NEXT: v_mov_b32_e32 v4, v19
-; GFX9-NEXT: v_mov_b32_e32 v8, v16
-; GFX9-NEXT: v_mov_b32_e32 v12, v17
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8bf16_to_v16i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v17, v3 :: v_dual_mov_b32 v16, v2
-; GFX11-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB54_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-NEXT: .LBB54_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB54_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_add_f32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v7, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v1
-; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v4
-; GFX11-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v18
-; GFX11-NEXT: v_bfe_u32 v6, v0, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v12, v7, v8 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 16, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v11, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v1, v11, v3, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_bfe_u32 v9, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v9, v9, v2, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v1, v4, vcc_lo
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v7 :: v_dual_lshlrev_b32 v1, 16, v16
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v3
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v9, v4, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v1
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v8
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v6, v8, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v9, v4, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_bfe_u32 v0, v1, 16, 1
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v8, v9, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v10, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v0, v0, v1, 0x7fff
-; GFX11-NEXT: v_perm_b32 v8, v6, v4, 0x7060302
-; GFX11-NEXT: v_add3_u32 v10, v10, v7, 0x7fff
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v10, v11, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_perm_b32 v1, v5, v12, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v0, v13, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v2, v3, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-NEXT: v_perm_b32 v7, v7, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: .LBB54_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v18
-; GFX11-NEXT: v_mov_b32_e32 v4, v19
-; GFX11-NEXT: v_mov_b32_e32 v8, v16
-; GFX11-NEXT: v_mov_b32_e32 v12, v17
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <8 x bfloat> %a1 to <16 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x bfloat> %a to <16 x i8>
- br label %end
-
-end:
- %phi = phi <16 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i8> %phi
-}
-
-define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i8_to_v8bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 24, v15
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 24, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v3, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v11, v1, v0
-; GCN-NEXT: v_or_b32_e32 v15, v18, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v19, v4
-; GCN-NEXT: v_or_b32_e32 v13, v6, v5
-; GCN-NEXT: v_or_b32_e32 v5, v20, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_or_b32_e32 v7, v21, v9
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: .LBB55_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v12
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v14
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v23, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v9
-; GCN-NEXT: v_or_b32_e32 v4, v22, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x300, v3
-; GCN-NEXT: v_or_b32_e32 v3, v21, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v7
-; GCN-NEXT: v_or_b32_e32 v7, v20, v8
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v6, v19, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_or_b32_e32 v1, v18, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v5
-; GCN-NEXT: v_or_b32_e32 v4, v6, v4
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x3000000, v0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v2
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; GCN-NEXT: .LBB55_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v11
-; GCN-NEXT: v_mov_b32_e32 v1, v15
-; GCN-NEXT: v_mov_b32_e32 v2, v17
-; GCN-NEXT: v_mov_b32_e32 v4, v13
-; GCN-NEXT: v_mov_b32_e32 v6, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i8_to_v8bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v18, v2
-; VI-NEXT: v_mov_b32_e32 v17, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB55_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB55_4
-; VI-NEXT: .LBB55_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB55_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr6
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB55_2
-; VI-NEXT: .LBB55_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v17
-; VI-NEXT: v_add_u16_e32 v1, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v4
-; VI-NEXT: v_add_u16_e32 v2, 3, v6
-; VI-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v4, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v4, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v12
-; VI-NEXT: v_add_u16_e32 v6, 3, v14
-; VI-NEXT: v_or_b32_sdwa v4, v7, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i8_to_v8bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v18, v2
-; GFX9-NEXT: v_mov_b32_e32 v17, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v7, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v5, 8, v15
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB55_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB55_4
-; GFX9-NEXT: .LBB55_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB55_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v18, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr6
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr7
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB55_2
-; GFX9-NEXT: .LBB55_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v17
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v21, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v4
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v9, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v7, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i8_to_v8bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v18, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v5, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v7, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v15
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB55_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB55_4
-; GFX11-NEXT: .LBB55_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB55_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v20
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v5
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v9
-; GFX11-NEXT: v_or_b32_e32 v6, v8, v11
-; GFX11-NEXT: v_or_b32_e32 v7, v10, v13
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr4
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB55_2
-; GFX11-NEXT: .LBB55_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v17, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v4, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v6, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v16, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v19, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v20, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v11, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v13, v10
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i8> %a, splat (i8 3)
- %a2 = bitcast <16 x i8> %a1 to <8 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i8> %a to <8 x bfloat>
- br label %end
-
-end:
- %phi = phi <8 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x bfloat> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.160bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.160bit.ll
deleted file mode 100644
index e7262375fbeb0..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.160bit.ll
+++ /dev/null
@@ -1,178 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <5 x float> @bitcast_v5i32_to_v5f32(<5 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i32_to_v5f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i32_to_v5f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i32_to_v5f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i32_to_v5f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v5
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i32> %a, splat (i32 3)
- %a2 = bitcast <5 x i32> %a1 to <5 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i32> %a to <5 x float>
- br label %end
-
-end:
- %phi = phi <5 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x float> %phi
-}
-
-define <5 x i32> @bitcast_v5f32_to_v5i32(<5 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f32_to_v5i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f32_to_v5i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f32_to_v5i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f32_to_v5i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v5
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <5 x float> %a1 to <5 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x float> %a to <5 x i32>
- br label %end
-
-end:
- %phi = phi <5 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
deleted file mode 100644
index 1185a12a474ea..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
+++ /dev/null
@@ -1,556 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define half @bitcast_i16_to_f16(i16 %a, i32 %b) {
-; GCN-LABEL: bitcast_i16_to_f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB0_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB0_4
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB0_3: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: .LBB0_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i16_to_f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i16_to_f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i16_to_f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v0, 3
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i16 %a, 3
- %a2 = bitcast i16 %a1 to half
- br label %end
-
-cmp.false:
- %a3 = bitcast i16 %a to half
- br label %end
-
-end:
- %phi = phi half [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret half %phi
-}
-
-define i16 @bitcast_f16_to_i16(half %a, i32 %b) {
-; GCN-LABEL: bitcast_f16_to_i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_mov_b32_e32 v0, 0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB1_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB1_4
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB1_3: ; %cmp.false
-; GCN-NEXT: v_mov_b32_e32 v0, v1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: .LBB1_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f16_to_i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f16_to_i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f16_e32 v0, 0x200, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f16_to_i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f16_e32 v0, 0x200, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd half %a, 0xH0200
- %a2 = bitcast half %a1 to i16
- br label %end
-
-cmp.false:
- %a3 = bitcast half %a to i16
- br label %end
-
-end:
- %phi = phi i16 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i16 %phi
-}
-
-define bfloat @bitcast_i16_to_bf16(i16 %a, i32 %b) {
-; GCN-LABEL: bitcast_i16_to_bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i16_to_bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i16_to_bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i16_to_bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v0, 3
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i16 %a, 3
- %a2 = bitcast i16 %a1 to bfloat
- br label %end
-
-cmp.false:
- %a3 = bitcast i16 %a to bfloat
- br label %end
-
-end:
- %phi = phi bfloat [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret bfloat %phi
-}
-
-define i16 @bitcast_bf16_to_i16(bfloat %a, i32 %b) {
-; GCN-LABEL: bitcast_bf16_to_i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_mov_b32_e32 v0, 0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_4
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB3_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: .LBB3_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_bf16_to_i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_bf16_to_i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_bf16_to_i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd bfloat %a, 0xR40C0
- %a2 = bitcast bfloat %a1 to i16
- br label %end
-
-cmp.false:
- %a3 = bitcast bfloat %a to i16
- br label %end
-
-end:
- %phi = phi i16 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i16 %phi
-}
-
-define bfloat @bitcast_f16_to_bf16(half %a, i32 %b) {
-; GCN-LABEL: bitcast_f16_to_bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB4_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB4_4
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB4_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: .LBB4_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f16_to_bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f16_to_bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f16_e32 v0, 0x200, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f16_to_bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f16_e32 v0, 0x200, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd half %a, 0xH0200
- %a2 = bitcast half %a1 to bfloat
- br label %end
-
-cmp.false:
- %a3 = bitcast half %a to bfloat
- br label %end
-
-end:
- %phi = phi bfloat [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret bfloat %phi
-}
-
-define half @bitcast_bf16_to_f16(bfloat %a, i32 %b) {
-; GCN-LABEL: bitcast_bf16_to_f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_4
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB5_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: .LBB5_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_bf16_to_f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_bfe_u32 v1, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v0
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
-; VI-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_bf16_to_f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
-; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_bf16_to_f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd bfloat %a, 0xR40C0
- %a2 = bitcast bfloat %a1 to half
- br label %end
-
-cmp.false:
- %a3 = bitcast bfloat %a to half
- br label %end
-
-end:
- %phi = phi half [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret half %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.192bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.192bit.ll
deleted file mode 100644
index 61f9232ea50a1..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.192bit.ll
+++ /dev/null
@@ -1,1068 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <6 x float> @bitcast_v6i32_to_v6f32(<6 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i32_to_v6f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i32_to_v6f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i32_to_v6f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i32_to_v6f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i32> %a, splat (i32 3)
- %a2 = bitcast <6 x i32> %a1 to <6 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i32> %a to <6 x float>
- br label %end
-
-end:
- %phi = phi <6 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x float> %phi
-}
-
-define <6 x i32> @bitcast_v6f32_to_v6i32(<6 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f32_to_v6i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f32_to_v6i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f32_to_v6i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f32_to_v6i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <6 x float> %a1 to <6 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x float> %a to <6 x i32>
- br label %end
-
-end:
- %phi = phi <6 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i32> %phi
-}
-
-define <3 x i64> @bitcast_v6i32_to_v3i64(<6 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i32_to_v3i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i32_to_v3i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i32_to_v3i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i32_to_v3i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i32> %a, splat (i32 3)
- %a2 = bitcast <6 x i32> %a1 to <3 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i32> %a to <3 x i64>
- br label %end
-
-end:
- %phi = phi <3 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x i64> %phi
-}
-
-define <6 x i32> @bitcast_v3i64_to_v6i32(<3 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3i64_to_v6i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3i64_to_v6i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3i64_to_v6i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3i64_to_v6i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <3 x i64> %a, splat (i64 3)
- %a2 = bitcast <3 x i64> %a1 to <6 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x i64> %a to <6 x i32>
- br label %end
-
-end:
- %phi = phi <6 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i32> %phi
-}
-
-define <3 x double> @bitcast_v6i32_to_v3f64(<6 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i32_to_v3f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i32_to_v3f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i32_to_v3f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i32_to_v3f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i32> %a, splat (i32 3)
- %a2 = bitcast <6 x i32> %a1 to <3 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i32> %a to <3 x double>
- br label %end
-
-end:
- %phi = phi <3 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x double> %phi
-}
-
-define <6 x i32> @bitcast_v3f64_to_v6i32(<3 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3f64_to_v6i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3f64_to_v6i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3f64_to_v6i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3f64_to_v6i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <3 x double> %a1 to <6 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x double> %a to <6 x i32>
- br label %end
-
-end:
- %phi = phi <6 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i32> %phi
-}
-
-define <3 x i64> @bitcast_v6f32_to_v3i64(<6 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f32_to_v3i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB6_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f32_to_v3i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f32_to_v3i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f32_to_v3i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <6 x float> %a1 to <3 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x float> %a to <3 x i64>
- br label %end
-
-end:
- %phi = phi <3 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x i64> %phi
-}
-
-define <6 x float> @bitcast_v3i64_to_v6f32(<3 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3i64_to_v6f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3i64_to_v6f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3i64_to_v6f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3i64_to_v6f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <3 x i64> %a, splat (i64 3)
- %a2 = bitcast <3 x i64> %a1 to <6 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x i64> %a to <6 x float>
- br label %end
-
-end:
- %phi = phi <6 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x float> %phi
-}
-
-define <3 x double> @bitcast_v6f32_to_v3f64(<6 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f32_to_v3f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f32_to_v3f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f32_to_v3f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f32_to_v3f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <6 x float> %a1 to <3 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x float> %a to <3 x double>
- br label %end
-
-end:
- %phi = phi <3 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x double> %phi
-}
-
-define <6 x float> @bitcast_v3f64_to_v6f32(<3 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3f64_to_v6f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3f64_to_v6f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3f64_to_v6f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB9_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB9_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3f64_to_v6f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <3 x double> %a1 to <6 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x double> %a to <6 x float>
- br label %end
-
-end:
- %phi = phi <6 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x float> %phi
-}
-
-define <3 x double> @bitcast_v3i64_to_v3f64(<3 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3i64_to_v3f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3i64_to_v3f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3i64_to_v3f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3i64_to_v3f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <3 x i64> %a, splat (i64 3)
- %a2 = bitcast <3 x i64> %a1 to <3 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x i64> %a to <3 x double>
- br label %end
-
-end:
- %phi = phi <3 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x double> %phi
-}
-
-define <3 x i64> @bitcast_v3f64_to_v3i64(<3 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3f64_to_v3i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3f64_to_v3i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3f64_to_v3i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3f64_to_v3i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v6
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <3 x double> %a1 to <3 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x double> %a to <3 x i64>
- br label %end
-
-end:
- %phi = phi <3 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x i64> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.224bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.224bit.ll
deleted file mode 100644
index 952be022750a6..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.224bit.ll
+++ /dev/null
@@ -1,194 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <7 x float> @bitcast_v7i32_to_v7f32(<7 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7i32_to_v7f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7i32_to_v7f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7i32_to_v7f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7i32_to_v7f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v7
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <7 x i32> %a, splat (i32 3)
- %a2 = bitcast <7 x i32> %a1 to <7 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x i32> %a to <7 x float>
- br label %end
-
-end:
- %phi = phi <7 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x float> %phi
-}
-
-define <7 x i32> @bitcast_v7f32_to_v7i32(<7 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7f32_to_v7i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7f32_to_v7i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7f32_to_v7i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7f32_to_v7i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v7
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <7 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <7 x float> %a1 to <7 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x float> %a to <7 x i32>
- br label %end
-
-end:
- %phi = phi <7 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
deleted file mode 100644
index 6e7b5dd33ea0b..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
+++ /dev/null
@@ -1,18360 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <8 x float> @bitcast_v8i32_to_v8f32(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <8 x i32> @bitcast_v8f32_to_v8i32(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <4 x i64> @bitcast_v8i32_to_v4i64(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <8 x i32> @bitcast_v4i64_to_v8i32(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <4 x double> @bitcast_v8i32_to_v4f64(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB4_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <8 x i32> @bitcast_v4f64_to_v8i32(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <16 x i16> @bitcast_v8i32_to_v16i16(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB6_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB6_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <8 x i32> @bitcast_v16i16_to_v8i32(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v4
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_4
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB7_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v0, v22
-; GCN-NEXT: v_or_b32_e32 v1, v1, v23
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v2, v2, v16
-; GCN-NEXT: v_or_b32_e32 v3, v3, v21
-; GCN-NEXT: v_or_b32_e32 v4, v4, v9
-; GCN-NEXT: v_or_b32_e32 v5, v5, v11
-; GCN-NEXT: v_or_b32_e32 v6, v6, v13
-; GCN-NEXT: v_or_b32_e32 v7, v7, v15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: .LBB7_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v2, v16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v21, v3
-; GCN-NEXT: v_or_b32_e32 v4, v9, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v13, v6
-; GCN-NEXT: v_or_b32_e32 v7, v15, v7
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 3
-; VI-NEXT: v_add_u16_e32 v8, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v8, v1
-; VI-NEXT: v_add_u16_e32 v8, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB7_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <16 x half> @bitcast_v8i32_to_v16f16(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v7
-; GCN-NEXT: v_mov_b32_e32 v18, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v21, v3
-; GCN-NEXT: v_mov_b32_e32 v22, v2
-; GCN-NEXT: v_mov_b32_e32 v23, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB8_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB8_4
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB8_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: .LBB8_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB8_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB8_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <8 x i32> @bitcast_v16f16_to_v8i32(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB9_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB9_4
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB9_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; GCN-NEXT: v_or_b32_e32 v0, v25, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v19, v2
-; GCN-NEXT: v_or_b32_e32 v3, v17, v3
-; GCN-NEXT: v_or_b32_e32 v4, v16, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v9, v6
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: .LBB9_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v11, v12
-; GCN-NEXT: v_or_b32_e32 v6, v9, v13
-; GCN-NEXT: v_or_b32_e32 v7, v8, v10
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 0x200
-; VI-NEXT: v_add_f16_sdwa v9, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v9
-; VI-NEXT: v_add_f16_sdwa v9, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v9
-; VI-NEXT: v_add_f16_sdwa v9, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v9
-; VI-NEXT: v_add_f16_sdwa v9, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v9
-; VI-NEXT: v_add_f16_sdwa v9, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v9
-; VI-NEXT: v_add_f16_sdwa v9, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v9
-; VI-NEXT: v_add_f16_sdwa v9, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v8, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v0, v0, v8
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <16 x bfloat> @bitcast_v8i32_to_v16bf16(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v23, v7
-; GCN-NEXT: v_mov_b32_e32 v22, v6
-; GCN-NEXT: v_mov_b32_e32 v21, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v19, v3
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_4
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB10_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v23
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v22
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v20
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: .LBB10_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v23
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB10_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB10_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <8 x i32> @bitcast_v16bf16_to_v8i32(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_4
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB11_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_alignbit_b32 v0, v0, v26, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v24, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v2, v20, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v18, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: .LBB11_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v13, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v9, v8, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v7, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v7
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v4
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v3
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v2
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v1
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s7
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v10, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v15, v6, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v7, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v12, v14 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v13, v10 :: v_dual_add_f32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v12, v10, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v11, v14 :: v_dual_lshlrev_b32 v11, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v9, 0x7060302
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v11 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add3_u32 v11, v12, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v14, v9, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v13, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v14, v9, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v12 :: v_dual_lshlrev_b32 v12, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v5, v5, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v4
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v11, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v11, v13, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v9
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v14, v10, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add3_u32 v13, v14, v10, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v11, v12 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_perm_b32 v4, v4, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v10
-; GFX11-NEXT: v_bfe_u32 v16, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v14, v11, 16, 1
-; GFX11-NEXT: v_perm_b32 v3, v3, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v12, v16, v2, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v13, v15 :: v_dual_lshlrev_b32 v15, 16, v0
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v12, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v13, v14, v11, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v11
-; GFX11-NEXT: v_bfe_u32 v15, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_bfe_u32 v16, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v12
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v13, v14, vcc_lo
-; GFX11-NEXT: v_add3_u32 v14, v15, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v16, v16, v12, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v14, v15 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_perm_b32 v1, v1, v11, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v16, v17, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v13, v13, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v13, v18, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v0, v12, 0x7060302
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <32 x i8> @bitcast_v8i32_to_v32i8(<8 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i32_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v28, v7
-; GCN-NEXT: v_mov_b32_e32 v24, v6
-; GCN-NEXT: v_mov_b32_e32 v20, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB12_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB12_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i32_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v5
-; VI-NEXT: v_mov_b32_e32 v32, v4
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v34, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: .LBB12_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_add_u32_e32 v35, vcc, 3, v35
-; VI-NEXT: v_add_u32_e32 v34, vcc, 3, v34
-; VI-NEXT: v_add_u32_e32 v33, vcc, 3, v33
-; VI-NEXT: v_add_u32_e32 v32, vcc, 3, v32
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB12_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v34
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v16, v32
-; VI-NEXT: v_mov_b32_e32 v20, v33
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v37
-; VI-NEXT: v_mov_b32_e32 v7, v36
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i32_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB12_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: v_add_u32_e32 v35, 3, v35
-; GFX9-NEXT: v_add_u32_e32 v34, 3, v34
-; GFX9-NEXT: v_add_u32_e32 v33, 3, v33
-; GFX9-NEXT: v_add_u32_e32 v32, 3, v32
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB12_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i32_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB12_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v39, 3, v39
-; GFX11-NEXT: v_add_nc_u32_e32 v37, 3, v37
-; GFX11-NEXT: v_add_nc_u32_e32 v35, 3, v35
-; GFX11-NEXT: v_add_nc_u32_e32 v33, 3, v33
-; GFX11-NEXT: v_add_nc_u32_e32 v32, 3, v32
-; GFX11-NEXT: v_add_nc_u32_e32 v34, 3, v34
-; GFX11-NEXT: v_add_nc_u32_e32 v36, 3, v36
-; GFX11-NEXT: v_add_nc_u32_e32 v38, 3, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB12_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i32> %a, splat (i32 3)
- %a2 = bitcast <8 x i32> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i32> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v8i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_4
-; GCN-NEXT: .LBB13_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB13_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v0, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v12, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: .LBB13_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v37, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v39, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v48, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v49, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v21, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v23, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v25, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x300, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x3000000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v8i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB13_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB13_4
-; VI-NEXT: .LBB13_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB13_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_2
-; VI-NEXT: .LBB13_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v7, 0x300
-; VI-NEXT: v_add_u16_e32 v2, 3, v33
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v8, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v13, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v8, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v28
-; VI-NEXT: v_add_u16_e32 v10, 3, v30
-; VI-NEXT: v_or_b32_sdwa v8, v11, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v7, v9, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v8i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB13_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB13_4
-; GFX9-NEXT: .LBB13_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB13_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_2
-; GFX9-NEXT: .LBB13_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v13, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v11, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v8i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_4
-; GFX11-NEXT: .LBB13_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB13_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v51
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v49
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v50
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v48
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v38
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v35
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v37
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v9, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v11, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v12, v16, v17
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB13_2
-; GFX11-NEXT: .LBB13_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v51, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v50, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v48, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v39, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v6, v18, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v3, v12, 3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v30, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v3, v35, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v37, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v19, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v21, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v12, v17, v16
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <8 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <8 x i32>
- br label %end
-
-end:
- %phi = phi <8 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i32> %phi
-}
-
-define <4 x i64> @bitcast_v8f32_to_v4i64(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <8 x float> @bitcast_v4i64_to_v8f32(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB15_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB15_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <4 x double> @bitcast_v8f32_to_v4f64(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <8 x float> @bitcast_v4f64_to_v8f32(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <16 x i16> @bitcast_v8f32_to_v16i16(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <8 x float> @bitcast_v16i16_to_v8f32(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v4
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_4
-; GCN-NEXT: .LBB19_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB19_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v0, v22
-; GCN-NEXT: v_or_b32_e32 v1, v1, v23
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v2, v2, v16
-; GCN-NEXT: v_or_b32_e32 v3, v3, v21
-; GCN-NEXT: v_or_b32_e32 v4, v4, v9
-; GCN-NEXT: v_or_b32_e32 v5, v5, v11
-; GCN-NEXT: v_or_b32_e32 v6, v6, v13
-; GCN-NEXT: v_or_b32_e32 v7, v7, v15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: .LBB19_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v2, v16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v21, v3
-; GCN-NEXT: v_or_b32_e32 v4, v9, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v13, v6
-; GCN-NEXT: v_or_b32_e32 v7, v15, v7
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 3
-; VI-NEXT: v_add_u16_e32 v8, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v8, v1
-; VI-NEXT: v_add_u16_e32 v8, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB19_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB19_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB19_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <16 x half> @bitcast_v8f32_to_v16f16(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v7
-; GCN-NEXT: v_mov_b32_e32 v18, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v21, v3
-; GCN-NEXT: v_mov_b32_e32 v22, v2
-; GCN-NEXT: v_mov_b32_e32 v23, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_4
-; GCN-NEXT: .LBB20_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB20_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: .LBB20_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <8 x float> @bitcast_v16f16_to_v8f32(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_4
-; GCN-NEXT: .LBB21_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB21_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; GCN-NEXT: v_or_b32_e32 v0, v25, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v19, v2
-; GCN-NEXT: v_or_b32_e32 v3, v17, v3
-; GCN-NEXT: v_or_b32_e32 v4, v16, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v9, v6
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: .LBB21_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v11, v12
-; GCN-NEXT: v_or_b32_e32 v6, v9, v13
-; GCN-NEXT: v_or_b32_e32 v7, v8, v10
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 0x200
-; VI-NEXT: v_add_f16_sdwa v9, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v9
-; VI-NEXT: v_add_f16_sdwa v9, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v9
-; VI-NEXT: v_add_f16_sdwa v9, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v9
-; VI-NEXT: v_add_f16_sdwa v9, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v9
-; VI-NEXT: v_add_f16_sdwa v9, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v9
-; VI-NEXT: v_add_f16_sdwa v9, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v9
-; VI-NEXT: v_add_f16_sdwa v9, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v8, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v0, v0, v8
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB21_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB21_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <16 x bfloat> @bitcast_v8f32_to_v16bf16(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v23, v7
-; GCN-NEXT: v_mov_b32_e32 v22, v6
-; GCN-NEXT: v_mov_b32_e32 v21, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v19, v3
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_4
-; GCN-NEXT: .LBB22_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB22_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v23
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v22
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v20
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: .LBB22_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v17
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v19
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v23
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <8 x float> @bitcast_v16bf16_to_v8f32(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_4
-; GCN-NEXT: .LBB23_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB23_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_alignbit_b32 v0, v0, v26, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v24, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v2, v20, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v18, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: .LBB23_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v13, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v9, v8, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB23_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v7, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v7
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v4
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v3
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v2
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v1
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB23_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB23_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s7
-; GFX9-NEXT: .LBB23_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB23_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v10, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v15, v6, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v7, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v12, v14 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v13, v10 :: v_dual_add_f32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v12, v10, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v11, v14 :: v_dual_lshlrev_b32 v11, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v9, 0x7060302
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v11 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add3_u32 v11, v12, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v14, v9, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v13, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v14, v9, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v12 :: v_dual_lshlrev_b32 v12, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v5, v5, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v4
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v11, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v11, v13, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v9
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v14, v10, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add3_u32 v13, v14, v10, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v11, v12 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_perm_b32 v4, v4, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v10
-; GFX11-NEXT: v_bfe_u32 v16, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v14, v11, 16, 1
-; GFX11-NEXT: v_perm_b32 v3, v3, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v12, v16, v2, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v13, v15 :: v_dual_lshlrev_b32 v15, 16, v0
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v12, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v13, v14, v11, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v11
-; GFX11-NEXT: v_bfe_u32 v15, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_bfe_u32 v16, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v12
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v13, v14, vcc_lo
-; GFX11-NEXT: v_add3_u32 v14, v15, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v16, v16, v12, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v14, v15 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_perm_b32 v1, v1, v11, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v16, v17, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v13, v13, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v13, v18, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v0, v12, 0x7060302
-; GFX11-NEXT: .LBB23_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <32 x i8> @bitcast_v8f32_to_v32i8(<8 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f32_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v28, v7
-; GCN-NEXT: v_mov_b32_e32 v24, v6
-; GCN-NEXT: v_mov_b32_e32 v20, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB24_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v32, 1.0, v32
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB24_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f32_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v5
-; VI-NEXT: v_mov_b32_e32 v32, v4
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v34, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: .LBB24_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: v_add_f32_e32 v35, 1.0, v35
-; VI-NEXT: v_add_f32_e32 v34, 1.0, v34
-; VI-NEXT: v_add_f32_e32 v33, 1.0, v33
-; VI-NEXT: v_add_f32_e32 v32, 1.0, v32
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB24_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v34
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v16, v32
-; VI-NEXT: v_mov_b32_e32 v20, v33
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v37
-; VI-NEXT: v_mov_b32_e32 v7, v36
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f32_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB24_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: v_add_f32_e32 v35, 1.0, v35
-; GFX9-NEXT: v_add_f32_e32 v34, 1.0, v34
-; GFX9-NEXT: v_add_f32_e32 v33, 1.0, v33
-; GFX9-NEXT: v_add_f32_e32 v32, 1.0, v32
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB24_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f32_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB24_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v39, 1.0, v39 :: v_dual_add_f32 v32, 1.0, v32
-; GFX11-NEXT: v_dual_add_f32 v37, 1.0, v37 :: v_dual_add_f32 v34, 1.0, v34
-; GFX11-NEXT: v_dual_add_f32 v35, 1.0, v35 :: v_dual_add_f32 v36, 1.0, v36
-; GFX11-NEXT: v_dual_add_f32 v33, 1.0, v33 :: v_dual_add_f32 v38, 1.0, v38
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB24_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <8 x float> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x float> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v8f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_4
-; GCN-NEXT: .LBB25_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB25_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v0, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v12, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: .LBB25_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v37, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v39, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v48, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v49, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v21, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v23, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v25, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x300, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x3000000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v8f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB25_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB25_4
-; VI-NEXT: .LBB25_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB25_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: .LBB25_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v7, 0x300
-; VI-NEXT: v_add_u16_e32 v2, 3, v33
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v8, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v13, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v8, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v28
-; VI-NEXT: v_add_u16_e32 v10, 3, v30
-; VI-NEXT: v_or_b32_sdwa v8, v11, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v7, v9, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v8f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB25_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB25_4
-; GFX9-NEXT: .LBB25_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB25_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: .LBB25_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v13, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v11, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v8f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_4
-; GFX11-NEXT: .LBB25_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB25_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v51
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v49
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v50
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v48
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v38
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v35
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v37
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v9, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v11, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v12, v16, v17
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: .LBB25_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v51, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v50, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v48, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v39, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v6, v18, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v3, v12, 3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v30, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v3, v35, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v37, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v19, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v21, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v12, v17, v16
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <8 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <8 x float>
- br label %end
-
-end:
- %phi = phi <8 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x float> %phi
-}
-
-define <4 x double> @bitcast_v4i64_to_v4f64(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: .LBB26_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB26_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: .LBB26_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <4 x i64> @bitcast_v4f64_to_v4i64(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: .LBB27_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: .LBB27_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: .LBB27_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: .LBB27_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <16 x i16> @bitcast_v4i64_to_v16i16(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB28_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <4 x i64> @bitcast_v16i16_to_v4i64(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v4
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_4
-; GCN-NEXT: .LBB29_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB29_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v0, v22
-; GCN-NEXT: v_or_b32_e32 v1, v1, v23
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v2, v2, v16
-; GCN-NEXT: v_or_b32_e32 v3, v3, v21
-; GCN-NEXT: v_or_b32_e32 v4, v4, v9
-; GCN-NEXT: v_or_b32_e32 v5, v5, v11
-; GCN-NEXT: v_or_b32_e32 v6, v6, v13
-; GCN-NEXT: v_or_b32_e32 v7, v7, v15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: .LBB29_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v2, v16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v21, v3
-; GCN-NEXT: v_or_b32_e32 v4, v9, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v13, v6
-; GCN-NEXT: v_or_b32_e32 v7, v15, v7
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 3
-; VI-NEXT: v_add_u16_e32 v8, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v8, v1
-; VI-NEXT: v_add_u16_e32 v8, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB29_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB29_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <16 x half> @bitcast_v4i64_to_v16f16(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v18, v7
-; GCN-NEXT: v_mov_b32_e32 v17, v6
-; GCN-NEXT: v_mov_b32_e32 v20, v5
-; GCN-NEXT: v_mov_b32_e32 v19, v4
-; GCN-NEXT: v_mov_b32_e32 v22, v3
-; GCN-NEXT: v_mov_b32_e32 v21, v2
-; GCN-NEXT: v_mov_b32_e32 v23, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_4
-; GCN-NEXT: .LBB30_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB30_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_2
-; GCN-NEXT: .LBB30_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v23, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v19
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v17
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v18, vcc
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB30_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB30_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <4 x i64> @bitcast_v16f16_to_v4i64(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_4
-; GCN-NEXT: .LBB31_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB31_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; GCN-NEXT: v_or_b32_e32 v0, v25, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v19, v2
-; GCN-NEXT: v_or_b32_e32 v3, v17, v3
-; GCN-NEXT: v_or_b32_e32 v4, v16, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v9, v6
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_2
-; GCN-NEXT: .LBB31_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v11, v12
-; GCN-NEXT: v_or_b32_e32 v6, v9, v13
-; GCN-NEXT: v_or_b32_e32 v7, v8, v10
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB31_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 0x200
-; VI-NEXT: v_add_f16_sdwa v9, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v9
-; VI-NEXT: v_add_f16_sdwa v9, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v9
-; VI-NEXT: v_add_f16_sdwa v9, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v9
-; VI-NEXT: v_add_f16_sdwa v9, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v9
-; VI-NEXT: v_add_f16_sdwa v9, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v9
-; VI-NEXT: v_add_f16_sdwa v9, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v9
-; VI-NEXT: v_add_f16_sdwa v9, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v8, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v0, v0, v8
-; VI-NEXT: .LBB31_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB31_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB31_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <16 x bfloat> @bitcast_v4i64_to_v16bf16(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v23, v7
-; GCN-NEXT: v_mov_b32_e32 v22, v6
-; GCN-NEXT: v_mov_b32_e32 v21, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v19, v3
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v1
-; GCN-NEXT: v_mov_b32_e32 v16, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_4
-; GCN-NEXT: .LBB32_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB32_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v23
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v22
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v20
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v16
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_2
-; GCN-NEXT: .LBB32_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v17, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v18
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v19, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v21, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v22
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v23, vcc
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB32_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB32_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <4 x i64> @bitcast_v16bf16_to_v4i64(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_4
-; GCN-NEXT: .LBB33_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB33_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_alignbit_b32 v0, v0, v26, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v24, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v2, v20, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v18, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_2
-; GCN-NEXT: .LBB33_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v13, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v9, v8, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB33_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v7, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v7
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v4
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v3
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v2
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v1
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB33_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB33_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s7
-; GFX9-NEXT: .LBB33_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB33_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v10, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v15, v6, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v7, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v12, v14 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v13, v10 :: v_dual_add_f32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v12, v10, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v11, v14 :: v_dual_lshlrev_b32 v11, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v9, 0x7060302
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v11 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add3_u32 v11, v12, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v14, v9, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v13, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v14, v9, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v12 :: v_dual_lshlrev_b32 v12, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v5, v5, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v4
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v11, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v11, v13, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v9
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v14, v10, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add3_u32 v13, v14, v10, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v11, v12 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_perm_b32 v4, v4, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v10
-; GFX11-NEXT: v_bfe_u32 v16, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v14, v11, 16, 1
-; GFX11-NEXT: v_perm_b32 v3, v3, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v12, v16, v2, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v13, v15 :: v_dual_lshlrev_b32 v15, 16, v0
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v12, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v13, v14, v11, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v11
-; GFX11-NEXT: v_bfe_u32 v15, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_bfe_u32 v16, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v12
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v13, v14, vcc_lo
-; GFX11-NEXT: v_add3_u32 v14, v15, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v16, v16, v12, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v14, v15 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_perm_b32 v1, v1, v11, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v16, v17, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v13, v13, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v13, v18, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v0, v12, 0x7060302
-; GFX11-NEXT: .LBB33_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <32 x i8> @bitcast_v4i64_to_v32i8(<4 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i64_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v28, v7
-; GCN-NEXT: v_mov_b32_e32 v24, v6
-; GCN-NEXT: v_mov_b32_e32 v20, v5
-; GCN-NEXT: v_mov_b32_e32 v16, v4
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB34_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_addc_u32_e32 v28, vcc, 0, v28, vcc
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v32, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v32, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v32, 8
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: .LBB34_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v8, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i64_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v5
-; VI-NEXT: v_mov_b32_e32 v32, v4
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v34, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: .LBB34_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v34, vcc, 3, v34
-; VI-NEXT: v_addc_u32_e32 v35, vcc, 0, v35, vcc
-; VI-NEXT: v_add_u32_e32 v32, vcc, 3, v32
-; VI-NEXT: v_addc_u32_e32 v33, vcc, 0, v33, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB34_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v34
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v16, v32
-; VI-NEXT: v_mov_b32_e32 v20, v33
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v37
-; VI-NEXT: v_mov_b32_e32 v7, v36
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i64_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB34_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v34, vcc, 3, v34
-; GFX9-NEXT: v_addc_co_u32_e32 v35, vcc, 0, v35, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v32, vcc, 3, v32
-; GFX9-NEXT: v_addc_co_u32_e32 v33, vcc, 0, v33, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB34_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i64_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB34_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v36, vcc_lo, v36, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v37, null, 0, v37, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v34, vcc_lo, v34, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v35, null, 0, v35, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v32, vcc_lo, v32, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v33, null, 0, v33, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v38, vcc_lo, v38, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v39, null, 0, v39, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB34_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i64> %a, splat (i64 3)
- %a2 = bitcast <4 x i64> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i64> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v4i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_4
-; GCN-NEXT: .LBB35_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB35_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v0, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v12, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_2
-; GCN-NEXT: .LBB35_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v37, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v39, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v48, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v49, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v21, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v23, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v25, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x300, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x3000000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v4i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_4
-; VI-NEXT: .LBB35_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB35_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_2
-; VI-NEXT: .LBB35_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v7, 0x300
-; VI-NEXT: v_add_u16_e32 v2, 3, v33
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v8, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v13, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v8, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v28
-; VI-NEXT: v_add_u16_e32 v10, 3, v30
-; VI-NEXT: v_or_b32_sdwa v8, v11, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v7, v9, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v4i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_4
-; GFX9-NEXT: .LBB35_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB35_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_2
-; GFX9-NEXT: .LBB35_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v13, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v11, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v4i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_4
-; GFX11-NEXT: .LBB35_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB35_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v51
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v49
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v50
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v48
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v38
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v35
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v37
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v9, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v11, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v12, v16, v17
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB35_2
-; GFX11-NEXT: .LBB35_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v51, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v50, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v48, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v39, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v6, v18, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v3, v12, 3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v30, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v3, v35, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v37, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v19, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v21, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v12, v17, v16
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <4 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <4 x i64>
- br label %end
-
-end:
- %phi = phi <4 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i64> %phi
-}
-
-define <16 x i16> @bitcast_v4f64_to_v16i16(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v7
-; GCN-NEXT: v_mov_b32_e32 v16, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v5
-; GCN-NEXT: v_mov_b32_e32 v18, v4
-; GCN-NEXT: v_mov_b32_e32 v21, v3
-; GCN-NEXT: v_mov_b32_e32 v20, v2
-; GCN-NEXT: v_mov_b32_e32 v23, v1
-; GCN-NEXT: v_mov_b32_e32 v22, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v13, v17, v16, 16
-; GCN-NEXT: v_alignbit_b32 v9, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v5, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v1, v23, v22, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v23
-; GCN-NEXT: .LBB36_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
-; GCN-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
-; GCN-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
-; GCN-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
-; GCN-NEXT: v_alignbit_b32 v13, v17, v16, 16
-; GCN-NEXT: v_alignbit_b32 v9, v19, v18, 16
-; GCN-NEXT: v_alignbit_b32 v5, v21, v20, 16
-; GCN-NEXT: v_alignbit_b32 v1, v23, v22, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v23
-; GCN-NEXT: .LBB36_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v22
-; GCN-NEXT: v_mov_b32_e32 v2, v23
-; GCN-NEXT: v_mov_b32_e32 v4, v20
-; GCN-NEXT: v_mov_b32_e32 v6, v21
-; GCN-NEXT: v_mov_b32_e32 v8, v18
-; GCN-NEXT: v_mov_b32_e32 v10, v19
-; GCN-NEXT: v_mov_b32_e32 v12, v16
-; GCN-NEXT: v_mov_b32_e32 v14, v17
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB36_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB36_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB36_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <4 x double> @bitcast_v16i16_to_v4f64(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v20, v6
-; GCN-NEXT: v_mov_b32_e32 v19, v4
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_4
-; GCN-NEXT: .LBB37_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB37_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v0, v22
-; GCN-NEXT: v_or_b32_e32 v1, v1, v23
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v2, v2, v16
-; GCN-NEXT: v_or_b32_e32 v3, v3, v21
-; GCN-NEXT: v_or_b32_e32 v4, v4, v9
-; GCN-NEXT: v_or_b32_e32 v5, v5, v11
-; GCN-NEXT: v_or_b32_e32 v6, v6, v13
-; GCN-NEXT: v_or_b32_e32 v7, v7, v15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_2
-; GCN-NEXT: .LBB37_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v17
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v22, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_or_b32_e32 v2, v16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v21, v3
-; GCN-NEXT: v_or_b32_e32 v4, v9, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v13, v6
-; GCN-NEXT: v_or_b32_e32 v7, v15, v7
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB37_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 3
-; VI-NEXT: v_add_u16_e32 v8, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: v_add_u16_e32 v8, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v8, v1
-; VI-NEXT: v_add_u16_e32 v8, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB37_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB37_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB37_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <16 x half> @bitcast_v4f64_to_v16f16(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: .LBB38_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: .LBB38_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v19
-; GCN-NEXT: v_mov_b32_e32 v1, v23
-; GCN-NEXT: v_mov_b32_e32 v2, v16
-; GCN-NEXT: v_mov_b32_e32 v3, v22
-; GCN-NEXT: v_mov_b32_e32 v4, v17
-; GCN-NEXT: v_mov_b32_e32 v5, v21
-; GCN-NEXT: v_mov_b32_e32 v6, v18
-; GCN-NEXT: v_mov_b32_e32 v7, v20
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB38_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB38_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB38_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB38_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB38_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB38_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <4 x double> @bitcast_v16f16_to_v4f64(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_4
-; GCN-NEXT: .LBB39_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB39_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; GCN-NEXT: v_or_b32_e32 v0, v25, v0
-; GCN-NEXT: v_or_b32_e32 v1, v23, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v19, v2
-; GCN-NEXT: v_or_b32_e32 v3, v17, v3
-; GCN-NEXT: v_or_b32_e32 v4, v16, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_or_b32_e32 v6, v9, v6
-; GCN-NEXT: v_or_b32_e32 v7, v8, v7
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_2
-; GCN-NEXT: .LBB39_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v11, v12
-; GCN-NEXT: v_or_b32_e32 v6, v9, v13
-; GCN-NEXT: v_or_b32_e32 v7, v8, v10
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB39_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 0x200
-; VI-NEXT: v_add_f16_sdwa v9, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v9
-; VI-NEXT: v_add_f16_sdwa v9, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v9
-; VI-NEXT: v_add_f16_sdwa v9, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v9
-; VI-NEXT: v_add_f16_sdwa v9, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v9
-; VI-NEXT: v_add_f16_sdwa v9, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v9
-; VI-NEXT: v_add_f16_sdwa v9, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v9
-; VI-NEXT: v_add_f16_sdwa v9, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v8, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v0, v0, v8
-; VI-NEXT: .LBB39_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB39_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB39_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <16 x bfloat> @bitcast_v4f64_to_v16bf16(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v3
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v2
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v1
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: .LBB40_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v3
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v2
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v1
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v0
-; GCN-NEXT: .LBB40_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v23
-; GCN-NEXT: v_mov_b32_e32 v1, v22
-; GCN-NEXT: v_mov_b32_e32 v2, v21
-; GCN-NEXT: v_mov_b32_e32 v3, v20
-; GCN-NEXT: v_mov_b32_e32 v4, v19
-; GCN-NEXT: v_mov_b32_e32 v5, v18
-; GCN-NEXT: v_mov_b32_e32 v6, v17
-; GCN-NEXT: v_mov_b32_e32 v7, v16
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB40_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB40_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB40_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB40_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB40_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB40_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <4 x double> @bitcast_v16bf16_to_v4f64(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v14
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_4
-; GCN-NEXT: .LBB41_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB41_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v23
-; GCN-NEXT: v_alignbit_b32 v0, v0, v26, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v24, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v2, v20, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v18, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_2
-; GCN-NEXT: .LBB41_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v23
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v6, v13, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v9, v8, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB41_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v7, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v7
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v4
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v3
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v2
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v1
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v8, 16
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB41_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB41_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v8, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s7
-; GFX9-NEXT: .LBB41_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB41_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v10, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v15, v6, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v7, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v12, v14 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v6
-; GFX11-NEXT: v_dual_cndmask_b32 v9, v13, v10 :: v_dual_add_f32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
-; GFX11-NEXT: v_bfe_u32 v12, v10, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v11, v14 :: v_dual_lshlrev_b32 v11, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v9, 0x7060302
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v11 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_add3_u32 v11, v12, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v14, v9, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v13, vcc_lo
-; GFX11-NEXT: v_add3_u32 v8, v14, v9, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v12 :: v_dual_lshlrev_b32 v12, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_perm_b32 v5, v5, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v4
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v11, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v11, v13, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v9
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v14, v10, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add3_u32 v13, v14, v10, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v3, v11, v12 :: v_dual_add_f32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_perm_b32 v4, v4, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v10
-; GFX11-NEXT: v_bfe_u32 v16, v2, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v14, v11, 16, 1
-; GFX11-NEXT: v_perm_b32 v3, v3, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add3_u32 v12, v16, v2, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v13, v15 :: v_dual_lshlrev_b32 v15, 16, v0
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v12, v13, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GFX11-NEXT: v_add3_u32 v13, v14, v11, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v11
-; GFX11-NEXT: v_bfe_u32 v15, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_bfe_u32 v16, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v12
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v13, v14, vcc_lo
-; GFX11-NEXT: v_add3_u32 v14, v15, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v16, v16, v12, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v14, v15 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_perm_b32 v1, v1, v11, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v16, v17, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v13, v13, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v13, v18, vcc_lo
-; GFX11-NEXT: v_perm_b32 v0, v0, v12, 0x7060302
-; GFX11-NEXT: .LBB41_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <32 x i8> @bitcast_v4f64_to_v32i8(<4 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4f64_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v35, v5
-; GCN-NEXT: v_mov_b32_e32 v34, v4
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v27, v7, v6, 24
-; GCN-NEXT: v_alignbit_b32 v26, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v25, v7, v6, 8
-; GCN-NEXT: v_alignbit_b32 v19, v35, v34, 24
-; GCN-NEXT: v_alignbit_b32 v18, v35, v34, 16
-; GCN-NEXT: v_alignbit_b32 v17, v35, v34, 8
-; GCN-NEXT: v_alignbit_b32 v11, v3, v2, 24
-; GCN-NEXT: v_alignbit_b32 v10, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v9, v3, v2, 8
-; GCN-NEXT: v_alignbit_b32 v38, v1, v0, 24
-; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v33, v1, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GCN-NEXT: .LBB42_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[34:35], v[34:35], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_alignbit_b32 v27, v7, v6, 24
-; GCN-NEXT: v_alignbit_b32 v26, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v25, v7, v6, 8
-; GCN-NEXT: v_alignbit_b32 v19, v35, v34, 24
-; GCN-NEXT: v_alignbit_b32 v18, v35, v34, 16
-; GCN-NEXT: v_alignbit_b32 v17, v35, v34, 8
-; GCN-NEXT: v_alignbit_b32 v11, v3, v2, 24
-; GCN-NEXT: v_alignbit_b32 v10, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v9, v3, v2, 8
-; GCN-NEXT: v_alignbit_b32 v38, v1, v0, 24
-; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v33, v1, v0, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GCN-NEXT: .LBB42_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_mov_b32_e32 v8, v2
-; GCN-NEXT: v_mov_b32_e32 v12, v3
-; GCN-NEXT: v_mov_b32_e32 v16, v34
-; GCN-NEXT: v_mov_b32_e32 v20, v35
-; GCN-NEXT: v_mov_b32_e32 v24, v6
-; GCN-NEXT: v_mov_b32_e32 v28, v7
-; GCN-NEXT: v_mov_b32_e32 v1, v33
-; GCN-NEXT: v_mov_b32_e32 v2, v32
-; GCN-NEXT: v_mov_b32_e32 v3, v38
-; GCN-NEXT: v_mov_b32_e32 v6, v37
-; GCN-NEXT: v_mov_b32_e32 v7, v36
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4f64_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v5
-; VI-NEXT: v_mov_b32_e32 v32, v4
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v34, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: .LBB42_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[32:33], v[32:33], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[34:35], v[34:35], 1.0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB42_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v34
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v16, v32
-; VI-NEXT: v_mov_b32_e32 v20, v33
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v37
-; VI-NEXT: v_mov_b32_e32 v7, v36
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4f64_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB42_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[32:33], v[32:33], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[34:35], v[34:35], 1.0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB42_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4f64_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB42_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[32:33], v[32:33], 1.0
-; GFX11-NEXT: v_add_f64 v[34:35], v[34:35], 1.0
-; GFX11-NEXT: v_add_f64 v[36:37], v[36:37], 1.0
-; GFX11-NEXT: v_add_f64 v[38:39], v[38:39], 1.0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB42_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <4 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <4 x double> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x double> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v4f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB43_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB43_4
-; GCN-NEXT: .LBB43_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB43_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v0, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v12, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_2
-; GCN-NEXT: .LBB43_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v37, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v38, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v39, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v48, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v49, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v21, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v16, v23, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v25, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v35, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v36, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v9, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v13, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v11, v15, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v13, v17, v18
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x300, v20
-; GCN-NEXT: v_or_b32_e32 v15, v19, v21
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 0x3000000, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v7
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v4f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB43_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB43_4
-; VI-NEXT: .LBB43_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB43_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB43_2
-; VI-NEXT: .LBB43_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v7, 0x300
-; VI-NEXT: v_add_u16_e32 v2, 3, v33
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v2, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v8
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v8, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v13, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v8, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v28
-; VI-NEXT: v_add_u16_e32 v10, 3, v30
-; VI-NEXT: v_or_b32_sdwa v8, v11, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v7, v9, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v4f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB43_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB43_4
-; GFX9-NEXT: .LBB43_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB43_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v35 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB43_2
-; GFX9-NEXT: .LBB43_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v8
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v38, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v19, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v17, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v15, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v13, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v11, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v4f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_4
-; GFX11-NEXT: .LBB43_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB43_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v51
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v49
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v50
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v48
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v38
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v35
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v37
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v9, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v11, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v12, v16, v17
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB43_2
-; GFX11-NEXT: .LBB43_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v8, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v51, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v50, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v48, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v39, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v6, v18, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_add_nc_u16 v3, v12, 3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v30, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_or_b32_e32 v3, v35, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v37, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v19, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v21, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v9, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v10, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v11, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v12, v17, v16
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v5, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v6, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v11, v12
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <4 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <4 x double>
- br label %end
-
-end:
- %phi = phi <4 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x double> %phi
-}
-
-define <16 x half> @bitcast_v16i16_to_v16f16(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v31, v15
-; GCN-NEXT: v_mov_b32_e32 v30, v14
-; GCN-NEXT: v_mov_b32_e32 v29, v13
-; GCN-NEXT: v_mov_b32_e32 v28, v12
-; GCN-NEXT: v_mov_b32_e32 v27, v11
-; GCN-NEXT: v_mov_b32_e32 v26, v10
-; GCN-NEXT: v_mov_b32_e32 v25, v9
-; GCN-NEXT: v_mov_b32_e32 v24, v8
-; GCN-NEXT: v_mov_b32_e32 v23, v7
-; GCN-NEXT: v_mov_b32_e32 v22, v6
-; GCN-NEXT: v_mov_b32_e32 v21, v5
-; GCN-NEXT: v_mov_b32_e32 v20, v4
-; GCN-NEXT: v_mov_b32_e32 v19, v3
-; GCN-NEXT: v_mov_b32_e32 v18, v2
-; GCN-NEXT: v_mov_b32_e32 v17, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB44_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB44_4
-; GCN-NEXT: .LBB44_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB44_3: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_2
-; GCN-NEXT: .LBB44_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v31
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v29
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB44_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 3
-; VI-NEXT: v_add_u16_sdwa v9, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v10, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v11, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v12, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v13, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v14, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v15, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v8, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_or_b32_e32 v6, v6, v15
-; VI-NEXT: v_or_b32_e32 v5, v5, v14
-; VI-NEXT: v_or_b32_e32 v4, v4, v13
-; VI-NEXT: v_or_b32_e32 v3, v3, v12
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_or_b32_e32 v1, v1, v10
-; VI-NEXT: v_or_b32_e32 v0, v0, v9
-; VI-NEXT: .LBB44_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB44_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB44_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <16 x i16> @bitcast_v16f16_to_v16i16(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB45_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_or_b32_e32 v14, v14, v16
-; GCN-NEXT: v_or_b32_e32 v10, v10, v17
-; GCN-NEXT: v_or_b32_e32 v6, v6, v18
-; GCN-NEXT: v_or_b32_e32 v2, v2, v19
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: .LBB45_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB45_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 0x200
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_or_b32_e32 v6, v15, v6
-; VI-NEXT: v_or_b32_e32 v5, v14, v5
-; VI-NEXT: v_or_b32_e32 v4, v13, v4
-; VI-NEXT: v_or_b32_e32 v3, v12, v3
-; VI-NEXT: v_or_b32_e32 v2, v11, v2
-; VI-NEXT: v_or_b32_e32 v1, v10, v1
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB45_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB45_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB45_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <16 x bfloat> @bitcast_v16i16_to_v16bf16(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v23, v14
-; GCN-NEXT: v_mov_b32_e32 v22, v12
-; GCN-NEXT: v_mov_b32_e32 v21, v10
-; GCN-NEXT: v_mov_b32_e32 v20, v8
-; GCN-NEXT: v_mov_b32_e32 v19, v6
-; GCN-NEXT: v_mov_b32_e32 v18, v4
-; GCN-NEXT: v_mov_b32_e32 v17, v2
-; GCN-NEXT: v_mov_b32_e32 v24, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_4
-; GCN-NEXT: .LBB46_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB46_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB46_2
-; GCN-NEXT: .LBB46_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v23
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v0, v15, v0
-; GCN-NEXT: v_or_b32_e32 v2, v13, v2
-; GCN-NEXT: v_or_b32_e32 v4, v11, v4
-; GCN-NEXT: v_or_b32_e32 v6, v9, v6
-; GCN-NEXT: v_or_b32_e32 v7, v7, v8
-; GCN-NEXT: v_or_b32_e32 v5, v5, v10
-; GCN-NEXT: v_or_b32_e32 v3, v3, v12
-; GCN-NEXT: v_or_b32_e32 v1, v1, v14
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB46_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v8, 3
-; VI-NEXT: v_add_u16_sdwa v9, v0, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v10, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v11, v2, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v12, v3, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v13, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v14, v5, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v15, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v8, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_or_b32_e32 v6, v6, v15
-; VI-NEXT: v_or_b32_e32 v5, v5, v14
-; VI-NEXT: v_or_b32_e32 v4, v4, v13
-; VI-NEXT: v_or_b32_e32 v3, v3, v12
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_or_b32_e32 v1, v1, v10
-; VI-NEXT: v_or_b32_e32 v0, v0, v9
-; VI-NEXT: .LBB46_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB46_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB46_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <16 x i16> @bitcast_v16bf16_to_v16i16(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB47_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB47_4
-; GCN-NEXT: .LBB47_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB47_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v22
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_2
-; GCN-NEXT: .LBB47_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v30
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v29
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v28
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v16
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v8
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v10
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v12
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_alignbit_b32 v0, v14, v0, 16
-; GCN-NEXT: v_alignbit_b32 v4, v19, v2, 16
-; GCN-NEXT: v_alignbit_b32 v8, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v12, v21, v5, 16
-; GCN-NEXT: v_alignbit_b32 v14, v15, v17, 16
-; GCN-NEXT: v_alignbit_b32 v10, v11, v9, 16
-; GCN-NEXT: v_alignbit_b32 v6, v7, v18, 16
-; GCN-NEXT: v_alignbit_b32 v2, v3, v13, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v24, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v23, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v22, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB47_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_bfe_u32 v10, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v10, vcc, v10, v9
-; VI-NEXT: v_add_u32_e32 v10, vcc, s6, v10
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v9, v10, v11, vcc
-; VI-NEXT: v_bfe_u32 v10, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v10, vcc, v10, v1
-; VI-NEXT: v_add_u32_e32 v10, vcc, s6, v10
-; VI-NEXT: v_or_b32_e32 v11, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v10, v11, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_bfe_u32 v11, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v11, vcc, v11, v10
-; VI-NEXT: v_add_u32_e32 v11, vcc, s6, v11
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc
-; VI-NEXT: v_bfe_u32 v11, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v11, vcc, v11, v2
-; VI-NEXT: v_add_u32_e32 v11, vcc, s6, v11
-; VI-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_bfe_u32 v12, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v12, vcc, v12, v11
-; VI-NEXT: v_add_u32_e32 v12, vcc, s6, v12
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v13, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v11, v12, v13, vcc
-; VI-NEXT: v_bfe_u32 v12, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v12, vcc, v12, v3
-; VI-NEXT: v_add_u32_e32 v12, vcc, s6, v12
-; VI-NEXT: v_or_b32_e32 v13, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v12, v13, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v4
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_bfe_u32 v13, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v13, vcc, v13, v12
-; VI-NEXT: v_add_u32_e32 v13, vcc, s6, v13
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v14, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v12, v13, v14, vcc
-; VI-NEXT: v_bfe_u32 v13, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v13, vcc, v13, v4
-; VI-NEXT: v_add_u32_e32 v13, vcc, s6, v13
-; VI-NEXT: v_or_b32_e32 v14, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v13, v14, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v5
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_bfe_u32 v14, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v14, vcc, v14, v13
-; VI-NEXT: v_add_u32_e32 v14, vcc, s6, v14
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v15, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v13, v14, v15, vcc
-; VI-NEXT: v_bfe_u32 v14, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v14, vcc, v14, v5
-; VI-NEXT: v_add_u32_e32 v14, vcc, s6, v14
-; VI-NEXT: v_or_b32_e32 v15, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v14, v15, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v6
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_bfe_u32 v15, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v15, vcc, v15, v14
-; VI-NEXT: v_add_u32_e32 v15, vcc, s6, v15
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v16, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
-; VI-NEXT: v_bfe_u32 v15, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v15, vcc, v15, v6
-; VI-NEXT: v_add_u32_e32 v15, vcc, s6, v15
-; VI-NEXT: v_or_b32_e32 v16, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v15, v16, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v7
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_bfe_u32 v16, v15, 16, 1
-; VI-NEXT: v_add_u32_e32 v16, vcc, v16, v15
-; VI-NEXT: v_add_u32_e32 v16, vcc, s6, v16
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; VI-NEXT: v_bfe_u32 v16, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v16, vcc, v16, v7
-; VI-NEXT: v_add_u32_e32 v16, vcc, 0x7fff, v16
-; VI-NEXT: v_or_b32_e32 v17, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v16, v17, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v15, 16
-; VI-NEXT: v_alignbit_b32 v6, v6, v14, 16
-; VI-NEXT: v_alignbit_b32 v5, v5, v13, 16
-; VI-NEXT: v_alignbit_b32 v4, v4, v12, 16
-; VI-NEXT: v_alignbit_b32 v3, v3, v11, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v10, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB47_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB47_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v10, v11, vcc
-; GFX9-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v10, v11, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_bfe_u32 v11, v10, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v11, v11, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc
-; GFX9-NEXT: v_bfe_u32 v11, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v11, v11, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_bfe_u32 v12, v11, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v12, v12, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v13, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v12, v13, vcc
-; GFX9-NEXT: v_bfe_u32 v12, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v12, v12, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v13, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v12, v13, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_bfe_u32 v13, v12, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v13, v13, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v14, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v13, v14, vcc
-; GFX9-NEXT: v_bfe_u32 v13, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v13, v13, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v14, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v13, v14, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_bfe_u32 v14, v13, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v14, v14, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v14, v15, vcc
-; GFX9-NEXT: v_bfe_u32 v14, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v14, v14, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v14, v15, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_bfe_u32 v15, v14, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v15, v15, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v15, v16, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v16, v17, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v15, s6
-; GFX9-NEXT: v_perm_b32 v6, v6, v14, s6
-; GFX9-NEXT: v_perm_b32 v5, v5, v13, s6
-; GFX9-NEXT: v_perm_b32 v4, v4, v12, s6
-; GFX9-NEXT: v_perm_b32 v3, v3, v11, s6
-; GFX9-NEXT: v_perm_b32 v2, v2, v10, s6
-; GFX9-NEXT: v_perm_b32 v1, v1, v9, s6
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s6
-; GFX9-NEXT: .LBB47_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB47_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v8, 16, v0
-; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v8, v11, v14 :: v_dual_and_b32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v14, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v12, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v0, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v0, v12, v15 :: v_dual_lshlrev_b32 v15, 16, v4
-; GFX11-NEXT: v_bfe_u32 v12, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v13, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v12, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v11, v12 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: v_perm_b32 v1, v1, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v10, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v11, v13, v10, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v2, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v13, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_bfe_u32 v14, v15, 16, 1
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v11, v11, v12 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_add3_u32 v14, v14, v15, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_add3_u32 v15, v18, v4, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v13, v13, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v14, v14, v17, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v19, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-NEXT: v_add3_u32 v18, v19, v12, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v6
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v15, v17 :: v_dual_add_f32 v15, 0x40c00000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_add3_u32 v17, v21, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v20, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
-; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v22, v6, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v21, v22, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
-; GFX11-NEXT: v_add3_u32 v23, v23, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v18
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v21, v22, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v15, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v23, v24, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v19, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v18, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_perm_b32 v4, v4, v14, 0x7060302
-; GFX11-NEXT: v_perm_b32 v5, v5, v12, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v13, v16, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v3, v3, v11, 0x7060302
-; GFX11-NEXT: .LBB47_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <32 x i8> @bitcast_v16i16_to_v32i8(<16 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i16_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v48, v15
-; GCN-NEXT: v_mov_b32_e32 v32, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v49, v11
-; GCN-NEXT: v_mov_b32_e32 v33, v10
-; GCN-NEXT: v_mov_b32_e32 v36, v8
-; GCN-NEXT: v_mov_b32_e32 v50, v7
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v38, v4
-; GCN-NEXT: v_mov_b32_e32 v51, v3
-; GCN-NEXT: v_mov_b32_e32 v35, v2
-; GCN-NEXT: v_mov_b32_e32 v39, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v50
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v9
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v49
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v48
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v39
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v32
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v51
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v50
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v49
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v48
-; GCN-NEXT: v_bfe_u32 v7, v51, 8, 8
-; GCN-NEXT: v_bfe_u32 v15, v50, 8, 8
-; GCN-NEXT: v_bfe_u32 v23, v49, 8, 8
-; GCN-NEXT: v_or_b32_e32 v0, v0, v52
-; GCN-NEXT: v_or_b32_e32 v4, v1, v53
-; GCN-NEXT: v_or_b32_e32 v8, v2, v54
-; GCN-NEXT: v_or_b32_e32 v12, v3, v55
-; GCN-NEXT: v_or_b32_e32 v16, v5, v40
-; GCN-NEXT: v_or_b32_e32 v20, v9, v41
-; GCN-NEXT: v_or_b32_e32 v24, v10, v42
-; GCN-NEXT: v_or_b32_e32 v28, v11, v43
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_bfe_u32 v31, v48, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: .LBB48_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v37
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v35
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v0, v42, v0
-; GCN-NEXT: v_or_b32_e32 v1, v43, v1
-; GCN-NEXT: v_or_b32_e32 v2, v40, v2
-; GCN-NEXT: v_or_b32_e32 v3, v41, v3
-; GCN-NEXT: v_or_b32_e32 v4, v54, v4
-; GCN-NEXT: v_or_b32_e32 v5, v55, v5
-; GCN-NEXT: v_or_b32_e32 v6, v52, v6
-; GCN-NEXT: v_or_b32_e32 v7, v53, v7
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v7
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: .LBB48_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i16_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v0
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v5
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v3
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[4:5]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[2:3]
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[0:1]
-; VI-NEXT: v_mov_b32_e32 v50, v0
-; VI-NEXT: v_mov_b32_e32 v48, v1
-; VI-NEXT: v_mov_b32_e32 v8, v2
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v16, v4
-; VI-NEXT: v_mov_b32_e32 v49, v5
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v51, v7
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: .LBB48_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 3
-; VI-NEXT: v_add_u16_sdwa v36, v1, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v32, v0, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v14, v3, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v10, v2, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v22, v5, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v4, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v30, v7, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v26, v6, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v48, 3, v1
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v36
-; VI-NEXT: v_add_u16_e32 v50, 3, v0
-; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; VI-NEXT: v_add_u16_e32 v35, 3, v3
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v14
-; VI-NEXT: v_add_u16_e32 v8, 3, v2
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; VI-NEXT: v_add_u16_e32 v49, 3, v5
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v22
-; VI-NEXT: v_add_u16_e32 v16, 3, v4
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v18
-; VI-NEXT: v_add_u16_e32 v51, 3, v7
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v30
-; VI-NEXT: v_add_u16_e32 v24, 3, v6
-; VI-NEXT: v_lshlrev_b32_e32 v6, 16, v26
-; VI-NEXT: v_or_b32_e32 v1, v48, v1
-; VI-NEXT: v_or_b32_e32 v0, v50, v0
-; VI-NEXT: v_or_b32_e32 v3, v35, v3
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_or_b32_e32 v5, v49, v5
-; VI-NEXT: v_or_b32_e32 v4, v16, v4
-; VI-NEXT: v_or_b32_e32 v7, v51, v7
-; VI-NEXT: v_or_b32_e32 v6, v24, v6
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[4:5]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[2:3]
-; VI-NEXT: v_lshrrev_b64 v[33:34], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_bfe_u32 v31, v30, 8, 8
-; VI-NEXT: v_bfe_u32 v23, v22, 8, 8
-; VI-NEXT: v_bfe_u32 v15, v14, 8, 8
-; VI-NEXT: v_bfe_u32 v39, v36, 8, 8
-; VI-NEXT: .LBB48_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v50
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v2, v32
-; VI-NEXT: v_mov_b32_e32 v3, v33
-; VI-NEXT: v_mov_b32_e32 v4, v48
-; VI-NEXT: v_mov_b32_e32 v5, v37
-; VI-NEXT: v_mov_b32_e32 v6, v36
-; VI-NEXT: v_mov_b32_e32 v7, v39
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v20, v49
-; VI-NEXT: v_mov_b32_e32 v28, v51
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i16_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB48_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v35, v35, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v34, v34, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v33, v33, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB48_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i16_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB48_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v39, v39, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v37, v37, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v35, v35, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v33, v33, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v34, v34, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v36, v36, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v38, v38, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB48_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i16> %a, splat (i16 3)
- %a2 = bitcast <16 x i16> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i16> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <16 x i16> @bitcast_v32i8_to_v16i16(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v16i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v17
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v25
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v24
-; GCN-NEXT: v_or_b32_e32 v1, v1, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v4, v4, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v7, v7, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v11, v11, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v0, v0, v51
-; GCN-NEXT: v_or_b32_e32 v8, v8, v52
-; GCN-NEXT: v_or_b32_e32 v14, v14, v54
-; GCN-NEXT: v_or_b32_e32 v15, v15, v55
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_or_b32_e32 v3, v32, v3
-; GCN-NEXT: v_or_b32_e32 v2, v33, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v34, v5
-; GCN-NEXT: v_or_b32_e32 v6, v35, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v9, v36, v9
-; GCN-NEXT: v_or_b32_e32 v10, v37, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v12, v39, v12
-; GCN-NEXT: v_or_b32_e32 v13, v38, v13
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v19, v1, v3
-; GCN-NEXT: v_or_b32_e32 v23, v4, v5
-; GCN-NEXT: v_or_b32_e32 v27, v7, v9
-; GCN-NEXT: v_or_b32_e32 v31, v11, v12
-; GCN-NEXT: v_or_b32_e32 v17, v0, v2
-; GCN-NEXT: v_or_b32_e32 v21, v8, v6
-; GCN-NEXT: v_or_b32_e32 v25, v14, v10
-; GCN-NEXT: v_or_b32_e32 v29, v15, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v9
-; GCN-NEXT: v_alignbit_b32 v1, v19, v2, 16
-; GCN-NEXT: v_alignbit_b32 v5, v23, v6, 16
-; GCN-NEXT: v_alignbit_b32 v9, v27, v10, 16
-; GCN-NEXT: v_alignbit_b32 v13, v31, v13, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v12
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB49_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v24
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v26
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v5, v53, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v9, v54, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v13, v50, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v52, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_or_b32_e32 v12, v49, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v0, v51, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
-; GCN-NEXT: v_or_b32_e32 v3, v38, v3
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_or_b32_e32 v7, v39, v7
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_or_b32_e32 v11, v37, v11
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_or_b32_e32 v15, v36, v15
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v10, v35, v10
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v14, v34, v14
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_or_b32_e32 v2, v33, v2
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v1, v3, v1
-; GCN-NEXT: v_or_b32_e32 v3, v7, v5
-; GCN-NEXT: v_or_b32_e32 v5, v11, v9
-; GCN-NEXT: v_or_b32_e32 v7, v15, v13
-; GCN-NEXT: v_or_b32_e32 v8, v10, v8
-; GCN-NEXT: v_or_b32_e32 v9, v14, v12
-; GCN-NEXT: v_or_b32_e32 v0, v2, v0
-; GCN-NEXT: v_or_b32_e32 v2, v6, v4
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v31, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v21, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v9
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v2
-; GCN-NEXT: v_alignbit_b32 v1, v19, v17, 16
-; GCN-NEXT: v_alignbit_b32 v5, v23, v21, 16
-; GCN-NEXT: v_alignbit_b32 v9, v27, v25, 16
-; GCN-NEXT: v_alignbit_b32 v13, v31, v29, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v31
-; GCN-NEXT: .LBB49_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v17
-; GCN-NEXT: v_mov_b32_e32 v2, v19
-; GCN-NEXT: v_mov_b32_e32 v4, v21
-; GCN-NEXT: v_mov_b32_e32 v6, v23
-; GCN-NEXT: v_mov_b32_e32 v8, v25
-; GCN-NEXT: v_mov_b32_e32 v10, v27
-; GCN-NEXT: v_mov_b32_e32 v12, v29
-; GCN-NEXT: v_mov_b32_e32 v14, v31
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v16i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v32, v6
-; VI-NEXT: v_mov_b32_e32 v34, v4
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB49_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB49_4
-; VI-NEXT: .LBB49_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB49_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v32, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_2
-; VI-NEXT: .LBB49_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v30
-; VI-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v1, 0x300
-; VI-NEXT: v_add_u16_sdwa v7, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v26
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v22
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v14
-; VI-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v12
-; VI-NEXT: v_or_b32_sdwa v12, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v10
-; VI-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v34
-; VI-NEXT: v_or_b32_sdwa v10, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v33
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v31
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v16
-; VI-NEXT: v_or_b32_sdwa v1, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v20
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v16i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v33, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v31, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB49_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB49_4
-; GFX9-NEXT: .LBB49_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB49_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v33, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v1, v2, v1, s6
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_2
-; GFX9-NEXT: .LBB49_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v12
-; GFX9-NEXT: v_or_b32_sdwa v0, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v0, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v0, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v35, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v10, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v9, v1, s6
-; GFX9-NEXT: v_perm_b32 v2, v8, v2, s6
-; GFX9-NEXT: v_perm_b32 v3, v11, v3, s6
-; GFX9-NEXT: v_perm_b32 v4, v15, v4, s6
-; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6
-; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6
-; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v16i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v34, v2 :: v_dual_mov_b32 v33, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v31, v6 :: v_dual_mov_b32 v32, v4
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_4
-; GFX11-NEXT: .LBB49_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB49_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v35
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v38
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v48
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v9
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v49
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v17
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v21
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v23
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v25
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v6, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v12, v11, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB49_2
-; GFX11-NEXT: .LBB49_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v30, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v20, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v23, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v25, v1
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v2, v21, v2
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v19, v3
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v4
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v17, v5
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v2
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v15, v0
-; GFX11-NEXT: v_add_nc_u16 v2, v8, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v39, v1
-; GFX11-NEXT: v_add_nc_u16 v14, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v13, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v11, v4
-; GFX11-NEXT: v_add_nc_u16 v4, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v48, v2
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v11, v37, v11
-; GFX11-NEXT: v_or_b32_e32 v13, v35, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v36, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_perm_b32 v0, v11, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v13, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v9, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v14, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v15, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v12, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v17, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <16 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <16 x i16>
- br label %end
-
-end:
- %phi = phi <16 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i16> %phi
-}
-
-define <16 x bfloat> @bitcast_v16f16_to_v16bf16(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB50_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB50_4
-; GCN-NEXT: .LBB50_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB50_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v31
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_2
-; GCN-NEXT: .LBB50_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v16
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v23
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB50_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v9, 0x200
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_or_b32_e32 v6, v15, v6
-; VI-NEXT: v_or_b32_e32 v5, v14, v5
-; VI-NEXT: v_or_b32_e32 v4, v13, v4
-; VI-NEXT: v_or_b32_e32 v3, v12, v3
-; VI-NEXT: v_or_b32_e32 v2, v11, v2
-; VI-NEXT: v_or_b32_e32 v1, v10, v1
-; VI-NEXT: v_or_b32_e32 v0, v8, v0
-; VI-NEXT: .LBB50_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB50_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB50_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
-
-define <16 x half> @bitcast_v16bf16_to_v16f16(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB51_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB51_4
-; GCN-NEXT: .LBB51_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB51_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_2
-; GCN-NEXT: .LBB51_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v30
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v29
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v28
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v27
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v26
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v24
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v20
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v16
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB51_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_bfe_u32 v9, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v8
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; VI-NEXT: v_bfe_u32 v9, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v9, vcc, v9, v0
-; VI-NEXT: v_add_u32_e32 v9, vcc, s6, v9
-; VI-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_bfe_u32 v10, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v10, vcc, v10, v9
-; VI-NEXT: v_add_u32_e32 v10, vcc, s6, v10
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v9, v10, v11, vcc
-; VI-NEXT: v_bfe_u32 v10, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v10, vcc, v10, v1
-; VI-NEXT: v_add_u32_e32 v10, vcc, s6, v10
-; VI-NEXT: v_or_b32_e32 v11, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v10, v11, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_bfe_u32 v11, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v11, vcc, v11, v10
-; VI-NEXT: v_add_u32_e32 v11, vcc, s6, v11
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc
-; VI-NEXT: v_bfe_u32 v11, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v11, vcc, v11, v2
-; VI-NEXT: v_add_u32_e32 v11, vcc, s6, v11
-; VI-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_bfe_u32 v12, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v12, vcc, v12, v11
-; VI-NEXT: v_add_u32_e32 v12, vcc, s6, v12
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v13, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v11, v12, v13, vcc
-; VI-NEXT: v_bfe_u32 v12, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v12, vcc, v12, v3
-; VI-NEXT: v_add_u32_e32 v12, vcc, s6, v12
-; VI-NEXT: v_or_b32_e32 v13, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v12, v13, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v4
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_bfe_u32 v13, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v13, vcc, v13, v12
-; VI-NEXT: v_add_u32_e32 v13, vcc, s6, v13
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v14, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v12, v13, v14, vcc
-; VI-NEXT: v_bfe_u32 v13, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v13, vcc, v13, v4
-; VI-NEXT: v_add_u32_e32 v13, vcc, s6, v13
-; VI-NEXT: v_or_b32_e32 v14, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v13, v14, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v5
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_bfe_u32 v14, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v14, vcc, v14, v13
-; VI-NEXT: v_add_u32_e32 v14, vcc, s6, v14
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v15, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v13, v14, v15, vcc
-; VI-NEXT: v_bfe_u32 v14, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v14, vcc, v14, v5
-; VI-NEXT: v_add_u32_e32 v14, vcc, s6, v14
-; VI-NEXT: v_or_b32_e32 v15, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v14, v15, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v6
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_bfe_u32 v15, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v15, vcc, v15, v14
-; VI-NEXT: v_add_u32_e32 v15, vcc, s6, v15
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v16, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
-; VI-NEXT: v_bfe_u32 v15, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v15, vcc, v15, v6
-; VI-NEXT: v_add_u32_e32 v15, vcc, s6, v15
-; VI-NEXT: v_or_b32_e32 v16, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v15, v16, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v7
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_bfe_u32 v16, v15, 16, 1
-; VI-NEXT: v_add_u32_e32 v16, vcc, v16, v15
-; VI-NEXT: v_add_u32_e32 v16, vcc, s6, v16
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; VI-NEXT: v_bfe_u32 v16, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v16, vcc, v16, v7
-; VI-NEXT: v_add_u32_e32 v16, vcc, 0x7fff, v16
-; VI-NEXT: v_or_b32_e32 v17, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v16, v17, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v15, 16
-; VI-NEXT: v_alignbit_b32 v6, v6, v14, 16
-; VI-NEXT: v_alignbit_b32 v5, v5, v13, 16
-; VI-NEXT: v_alignbit_b32 v4, v4, v12, 16
-; VI-NEXT: v_alignbit_b32 v3, v3, v11, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v10, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v9, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v8, 16
-; VI-NEXT: .LBB51_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB51_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v9, v9, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
-; GFX9-NEXT: v_bfe_u32 v9, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v10, v11, vcc
-; GFX9-NEXT: v_bfe_u32 v10, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v10, v11, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_bfe_u32 v11, v10, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v11, v11, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc
-; GFX9-NEXT: v_bfe_u32 v11, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v11, v11, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_bfe_u32 v12, v11, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v12, v12, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v13, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v12, v13, vcc
-; GFX9-NEXT: v_bfe_u32 v12, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v12, v12, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v13, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v12, v13, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_bfe_u32 v13, v12, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v13, v13, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v14, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v13, v14, vcc
-; GFX9-NEXT: v_bfe_u32 v13, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v13, v13, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v14, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v13, v14, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_bfe_u32 v14, v13, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v14, v14, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v14, v15, vcc
-; GFX9-NEXT: v_bfe_u32 v14, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v14, v14, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v14, v15, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_bfe_u32 v15, v14, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v15, v15, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v15, v16, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v16, v17, vcc
-; GFX9-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v16, v16, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v16, v17, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v7, v7, v15, s6
-; GFX9-NEXT: v_perm_b32 v6, v6, v14, s6
-; GFX9-NEXT: v_perm_b32 v5, v5, v13, s6
-; GFX9-NEXT: v_perm_b32 v4, v4, v12, s6
-; GFX9-NEXT: v_perm_b32 v3, v3, v11, s6
-; GFX9-NEXT: v_perm_b32 v2, v2, v10, s6
-; GFX9-NEXT: v_perm_b32 v1, v1, v9, s6
-; GFX9-NEXT: v_perm_b32 v0, v0, v8, s6
-; GFX9-NEXT: .LBB51_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v8
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB51_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v8, 16, v0
-; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v13, v9, 16, 1
-; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v13, v13, v9, 0x7fff
-; GFX11-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v8, v11, v14 :: v_dual_and_b32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v14, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v12, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v12, v12, v0, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v0, v12, v15 :: v_dual_lshlrev_b32 v15, 16, v4
-; GFX11-NEXT: v_bfe_u32 v12, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v13, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v12, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v11, v12 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: v_perm_b32 v1, v1, v9, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v13, v10, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_add3_u32 v11, v13, v10, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v2, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v13, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc_lo
-; GFX11-NEXT: v_add3_u32 v11, v13, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_bfe_u32 v14, v15, 16, 1
-; GFX11-NEXT: v_bfe_u32 v13, v3, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v11, v11, v12 :: v_dual_lshlrev_b32 v12, 16, v5
-; GFX11-NEXT: v_add3_u32 v14, v14, v15, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_add3_u32 v15, v18, v4, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v13, v13, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v14, v14, v17, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v19, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-NEXT: v_add3_u32 v18, v19, v12, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v6
-; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v4, v15, v17 :: v_dual_add_f32 v15, 0x40c00000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_add3_u32 v17, v21, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v20, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
-; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v22, v6, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v21, v22, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
-; GFX11-NEXT: v_add3_u32 v23, v23, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v18
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v21, v22, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v15, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v23, v24, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v19, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v7, v7, v18, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_perm_b32 v4, v4, v14, 0x7060302
-; GFX11-NEXT: v_perm_b32 v5, v5, v12, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v13, v16, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v3, v3, v11, 0x7060302
-; GFX11-NEXT: .LBB51_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <32 x i8> @bitcast_v16f16_to_v32i8(<16 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f16_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v17, v14
-; GCN-NEXT: v_mov_b32_e32 v18, v6
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v17
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB52_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB52_4
-; GCN-NEXT: .LBB52_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB52_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v30
-; GCN-NEXT: v_bfe_u32 v7, v6, 8, 8
-; GCN-NEXT: v_bfe_u32 v15, v14, 8, 8
-; GCN-NEXT: v_bfe_u32 v23, v22, 8, 8
-; GCN-NEXT: v_or_b32_e32 v0, v33, v0
-; GCN-NEXT: v_or_b32_e32 v4, v32, v1
-; GCN-NEXT: v_or_b32_e32 v8, v35, v2
-; GCN-NEXT: v_or_b32_e32 v12, v34, v3
-; GCN-NEXT: v_or_b32_e32 v16, v38, v5
-; GCN-NEXT: v_or_b32_e32 v20, v36, v9
-; GCN-NEXT: v_or_b32_e32 v24, v49, v10
-; GCN-NEXT: v_or_b32_e32 v28, v48, v11
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_bfe_u32 v31, v30, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB52_2
-; GCN-NEXT: .LBB52_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v32
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v6
-; GCN-NEXT: v_bfe_u32 v7, v6, 8, 8
-; GCN-NEXT: v_bfe_u32 v15, v14, 8, 8
-; GCN-NEXT: v_bfe_u32 v23, v22, 8, 8
-; GCN-NEXT: v_or_b32_e32 v24, v1, v0
-; GCN-NEXT: v_or_b32_e32 v28, v2, v12
-; GCN-NEXT: v_or_b32_e32 v16, v4, v3
-; GCN-NEXT: v_or_b32_e32 v20, v5, v17
-; GCN-NEXT: v_or_b32_e32 v8, v8, v18
-; GCN-NEXT: v_or_b32_e32 v12, v9, v19
-; GCN-NEXT: v_or_b32_e32 v0, v11, v10
-; GCN-NEXT: v_or_b32_e32 v4, v13, v21
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_bfe_u32 v31, v30, 8, 8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f16_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v35, v5
-; VI-NEXT: v_mov_b32_e32 v34, v4
-; VI-NEXT: v_mov_b32_e32 v33, v3
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB52_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB52_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB52_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v5, 0x200
-; VI-NEXT: v_add_f16_sdwa v14, v33, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v36, v1, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v14
-; VI-NEXT: v_add_f16_e32 v33, 0x200, v33
-; VI-NEXT: v_add_f16_sdwa v10, v32, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v36
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_or_b32_e32 v12, v33, v8
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v10
-; VI-NEXT: v_add_f16_e32 v32, 0x200, v32
-; VI-NEXT: v_add_f16_sdwa v22, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v1, v2
-; VI-NEXT: v_add_f16_sdwa v2, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v32, v8
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v22
-; VI-NEXT: v_add_f16_e32 v35, 0x200, v35
-; VI-NEXT: v_add_f16_sdwa v18, v34, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v30, v7, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v26, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v2
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v9, v35, v8
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v18
-; VI-NEXT: v_add_f16_e32 v34, 0x200, v34
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v30
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v26
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v3, v0, v3
-; VI-NEXT: v_or_b32_e32 v8, v34, v8
-; VI-NEXT: v_or_b32_e32 v16, v7, v13
-; VI-NEXT: v_or_b32_e32 v15, v6, v5
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[8:9]
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v11
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v8
-; VI-NEXT: v_bfe_u32 v31, v30, 8, 8
-; VI-NEXT: v_bfe_u32 v23, v22, 8, 8
-; VI-NEXT: v_bfe_u32 v15, v14, 8, 8
-; VI-NEXT: v_bfe_u32 v37, v36, 8, 8
-; VI-NEXT: .LBB52_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v32
-; VI-NEXT: v_mov_b32_e32 v12, v33
-; VI-NEXT: v_mov_b32_e32 v16, v34
-; VI-NEXT: v_mov_b32_e32 v20, v35
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v36
-; VI-NEXT: v_mov_b32_e32 v7, v37
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f16_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB52_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB52_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB52_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v35, v35, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v34, v34, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v33, v33, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v32, v32, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: .LBB52_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f16_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB52_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB52_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB52_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v39, 0x200, v39 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v37, 0x200, v37 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v35, 0x200, v35 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v33, 0x200, v33 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v34, 0x200, v34 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v36, 0x200, v36 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v38, 0x200, v38 op_sel_hi:[0,1]
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB52_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <16 x half> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x half> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <16 x half> @bitcast_v32i8_to_v16f16(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v16f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v0, v32
-; GCN-NEXT: v_or_b32_e32 v1, v1, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v34
-; GCN-NEXT: v_or_b32_e32 v3, v3, v35
-; GCN-NEXT: v_or_b32_e32 v4, v4, v36
-; GCN-NEXT: v_or_b32_e32 v5, v5, v37
-; GCN-NEXT: v_or_b32_e32 v6, v6, v38
-; GCN-NEXT: v_or_b32_e32 v7, v7, v39
-; GCN-NEXT: v_or_b32_e32 v8, v8, v48
-; GCN-NEXT: v_or_b32_e32 v9, v9, v49
-; GCN-NEXT: v_or_b32_e32 v10, v10, v50
-; GCN-NEXT: v_or_b32_e32 v11, v11, v51
-; GCN-NEXT: v_or_b32_e32 v12, v12, v52
-; GCN-NEXT: v_or_b32_e32 v13, v13, v53
-; GCN-NEXT: v_or_b32_e32 v14, v14, v54
-; GCN-NEXT: v_or_b32_e32 v15, v15, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB53_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB53_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v30
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_or_b32_e32 v3, v54, v3
-; GCN-NEXT: v_or_b32_e32 v5, v53, v5
-; GCN-NEXT: v_or_b32_e32 v7, v52, v7
-; GCN-NEXT: v_or_b32_e32 v9, v51, v9
-; GCN-NEXT: v_or_b32_e32 v11, v50, v11
-; GCN-NEXT: v_or_b32_e32 v13, v49, v13
-; GCN-NEXT: v_or_b32_e32 v15, v48, v15
-; GCN-NEXT: v_or_b32_e32 v14, v39, v14
-; GCN-NEXT: v_or_b32_e32 v12, v38, v12
-; GCN-NEXT: v_or_b32_e32 v10, v37, v10
-; GCN-NEXT: v_or_b32_e32 v8, v36, v8
-; GCN-NEXT: v_or_b32_e32 v6, v35, v6
-; GCN-NEXT: v_or_b32_e32 v4, v34, v4
-; GCN-NEXT: v_or_b32_e32 v2, v33, v2
-; GCN-NEXT: v_or_b32_e32 v0, v32, v0
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 0x300, v1
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v16
-; GCN-NEXT: .LBB53_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v31
-; GCN-NEXT: v_mov_b32_e32 v2, v23
-; GCN-NEXT: v_mov_b32_e32 v4, v19
-; GCN-NEXT: v_mov_b32_e32 v6, v27
-; GCN-NEXT: v_mov_b32_e32 v8, v17
-; GCN-NEXT: v_mov_b32_e32 v10, v21
-; GCN-NEXT: v_mov_b32_e32 v12, v25
-; GCN-NEXT: v_mov_b32_e32 v14, v29
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v16f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v32, v6
-; VI-NEXT: v_mov_b32_e32 v34, v4
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB53_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB53_4
-; VI-NEXT: .LBB53_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB53_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v32, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB53_2
-; VI-NEXT: .LBB53_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v30
-; VI-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v1, 0x300
-; VI-NEXT: v_add_u16_sdwa v7, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v26
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v22
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v14
-; VI-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v12
-; VI-NEXT: v_or_b32_sdwa v12, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v10
-; VI-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v34
-; VI-NEXT: v_or_b32_sdwa v10, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v33
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v31
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v16
-; VI-NEXT: v_or_b32_sdwa v1, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v20
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v16f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v33, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v31, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB53_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB53_4
-; GFX9-NEXT: .LBB53_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB53_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v33, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v1, v2, v1, s6
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB53_2
-; GFX9-NEXT: .LBB53_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v12
-; GFX9-NEXT: v_or_b32_sdwa v0, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v0, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v0, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v35, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v10, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v9, v1, s6
-; GFX9-NEXT: v_perm_b32 v2, v8, v2, s6
-; GFX9-NEXT: v_perm_b32 v3, v11, v3, s6
-; GFX9-NEXT: v_perm_b32 v4, v15, v4, s6
-; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6
-; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6
-; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v16f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v34, v2 :: v_dual_mov_b32 v33, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v31, v6 :: v_dual_mov_b32 v32, v4
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB53_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB53_4
-; GFX11-NEXT: .LBB53_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB53_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v35
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v38
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v48
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v9
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v49
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v17
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v21
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v23
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v25
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v6, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v12, v11, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB53_2
-; GFX11-NEXT: .LBB53_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v30, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v20, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v23, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v25, v1
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v2, v21, v2
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v19, v3
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v4
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v17, v5
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v2
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v15, v0
-; GFX11-NEXT: v_add_nc_u16 v2, v8, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v39, v1
-; GFX11-NEXT: v_add_nc_u16 v14, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v13, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v11, v4
-; GFX11-NEXT: v_add_nc_u16 v4, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v48, v2
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v11, v37, v11
-; GFX11-NEXT: v_or_b32_e32 v13, v35, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v36, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_perm_b32 v0, v11, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v13, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v9, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v14, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v15, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v12, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v17, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <16 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <16 x half>
- br label %end
-
-end:
- %phi = phi <16 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x half> %phi
-}
-
-define <32 x i8> @bitcast_v16bf16_to_v32i8(<16 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16bf16_to_v32i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v14
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB54_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB54_4
-; GCN-NEXT: .LBB54_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB54_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v53
-; GCN-NEXT: v_alignbit_b32 v0, v0, v33, 16
-; GCN-NEXT: v_alignbit_b32 v4, v6, v32, 16
-; GCN-NEXT: v_alignbit_b32 v8, v1, v36, 16
-; GCN-NEXT: v_alignbit_b32 v12, v14, v35, 16
-; GCN-NEXT: v_alignbit_b32 v16, v2, v48, 16
-; GCN-NEXT: v_alignbit_b32 v20, v22, v39, 16
-; GCN-NEXT: v_alignbit_b32 v24, v3, v52, 16
-; GCN-NEXT: v_alignbit_b32 v28, v30, v51, 16
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB54_2
-; GCN-NEXT: .LBB54_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v34
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GCN-NEXT: v_alignbit_b32 v24, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v28, v30, v2, 16
-; GCN-NEXT: v_alignbit_b32 v16, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v20, v22, v5, 16
-; GCN-NEXT: v_alignbit_b32 v8, v17, v8, 16
-; GCN-NEXT: v_alignbit_b32 v12, v14, v9, 16
-; GCN-NEXT: v_alignbit_b32 v0, v18, v11, 16
-; GCN-NEXT: v_alignbit_b32 v4, v6, v13, 16
-; GCN-NEXT: v_alignbit_b32 v3, v4, v0, 24
-; GCN-NEXT: v_alignbit_b32 v2, v4, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v4, v0, 8
-; GCN-NEXT: v_alignbit_b32 v11, v12, v8, 24
-; GCN-NEXT: v_alignbit_b32 v10, v12, v8, 16
-; GCN-NEXT: v_alignbit_b32 v9, v12, v8, 8
-; GCN-NEXT: v_alignbit_b32 v19, v20, v16, 24
-; GCN-NEXT: v_alignbit_b32 v18, v20, v16, 16
-; GCN-NEXT: v_alignbit_b32 v17, v20, v16, 8
-; GCN-NEXT: v_alignbit_b32 v27, v28, v24, 24
-; GCN-NEXT: v_alignbit_b32 v26, v28, v24, 16
-; GCN-NEXT: v_alignbit_b32 v25, v28, v24, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 8, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v31
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16bf16_to_v32i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v5
-; VI-NEXT: v_mov_b32_e32 v32, v4
-; VI-NEXT: v_mov_b32_e32 v35, v3
-; VI-NEXT: v_mov_b32_e32 v34, v2
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB54_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: .LBB54_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB54_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_bfe_u32 v3, v1, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v1
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_bfe_u32 v3, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v0
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v35
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v35
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v35, v3, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v34
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v34
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v34, v3, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v33
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v33, v3, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v32
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v32, v3, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v7
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v7, v3, v2, 16
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_bfe_u32 v4, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v4, vcc, v4, v3
-; VI-NEXT: v_add_u32_e32 v4, vcc, s6, v4
-; VI-NEXT: v_or_b32_e32 v5, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v6, v3, v2, 16
-; VI-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; VI-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; VI-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; VI-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; VI-NEXT: .LBB54_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v8, v34
-; VI-NEXT: v_mov_b32_e32 v12, v35
-; VI-NEXT: v_mov_b32_e32 v16, v32
-; VI-NEXT: v_mov_b32_e32 v20, v33
-; VI-NEXT: v_mov_b32_e32 v24, v6
-; VI-NEXT: v_mov_b32_e32 v28, v7
-; VI-NEXT: v_mov_b32_e32 v1, v38
-; VI-NEXT: v_mov_b32_e32 v6, v37
-; VI-NEXT: v_mov_b32_e32 v7, v36
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16bf16_to_v32i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v33, v5
-; GFX9-NEXT: v_mov_b32_e32 v32, v4
-; GFX9-NEXT: v_mov_b32_e32 v35, v3
-; GFX9-NEXT: v_mov_b32_e32 v34, v2
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB54_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v0
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[6:7]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[32:33]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[34:35]
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX9-NEXT: .LBB54_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB54_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v3, v3, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v3, v3, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v3, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_add3_u32 v3, v3, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v3, v5, vcc
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
-; GFX9-NEXT: v_perm_b32 v3, v0, v5, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v35
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v1, v1, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v9, vcc
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v35
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v9, v9, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v34
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v34
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v10, v11, vcc
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v11, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v10, v11, vcc
-; GFX9-NEXT: v_perm_b32 v11, v9, v13, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v33
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v14, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v33
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v10, v14, vcc
-; GFX9-NEXT: v_bfe_u32 v10, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v10, v10, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v15, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v10, v15, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v32
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v15, v16, vcc
-; GFX9-NEXT: v_bfe_u32 v15, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v15, v15, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v15, v16, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_bfe_u32 v15, v7, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v16, v19, vcc
-; GFX9-NEXT: v_add3_u32 v15, v15, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v16, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_bfe_u32 v16, v15, 16, 1
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_add3_u32 v16, v16, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_bfe_u32 v15, v6, 16, 1
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v16, v20, vcc
-; GFX9-NEXT: v_add3_u32 v15, v15, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v16, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v15, v16, vcc
-; GFX9-NEXT: v_perm_b32 v4, v8, v2, s7
-; GFX9-NEXT: v_perm_b32 v12, v1, v0, s7
-; GFX9-NEXT: v_perm_b32 v10, v17, v14, s7
-; GFX9-NEXT: v_perm_b32 v9, v9, v18, s7
-; GFX9-NEXT: v_perm_b32 v16, v7, v19, s7
-; GFX9-NEXT: v_perm_b32 v15, v6, v20, s7
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v19
-; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v20
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v18
-; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 8, v15
-; GFX9-NEXT: v_lshrrev_b64 v[27:28], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v9
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v9, 8, v11
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v5, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v3
-; GFX9-NEXT: v_lshrrev_b64 v[3:4], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v8
-; GFX9-NEXT: .LBB54_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-NEXT: v_mov_b32_e32 v8, v34
-; GFX9-NEXT: v_mov_b32_e32 v12, v35
-; GFX9-NEXT: v_mov_b32_e32 v16, v32
-; GFX9-NEXT: v_mov_b32_e32 v20, v33
-; GFX9-NEXT: v_mov_b32_e32 v24, v6
-; GFX9-NEXT: v_mov_b32_e32 v28, v7
-; GFX9-NEXT: v_mov_b32_e32 v1, v38
-; GFX9-NEXT: v_mov_b32_e32 v6, v37
-; GFX9-NEXT: v_mov_b32_e32 v7, v36
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16bf16_to_v32i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v7 :: v_dual_mov_b32 v32, v6
-; GFX11-NEXT: v_dual_mov_b32 v35, v5 :: v_dual_mov_b32 v34, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v3 :: v_dual_mov_b32 v36, v2
-; GFX11-NEXT: v_dual_mov_b32 v39, v1 :: v_dual_mov_b32 v38, v0
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: ; implicit-def: $vgpr6
-; GFX11-NEXT: ; implicit-def: $vgpr7
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB54_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-NEXT: .LBB54_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB54_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v39
-; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v39
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v38
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v38
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v36
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v34
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v6, v1, 16, 1
-; GFX11-NEXT: v_bfe_u32 v4, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v1
-; GFX11-NEXT: v_add3_u32 v6, v6, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add3_u32 v1, v9, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v4
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v0, 0x400000, v2
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v37
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v0, vcc_lo
-; GFX11-NEXT: v_dual_add_f32 v7, 0x40c00000, v8 :: v_dual_add_f32 v6, 0x40c00000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v36
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v7
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v9, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v9, v7, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v9, v9, v7, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v12, v3, 16, 1
-; GFX11-NEXT: v_perm_b32 v1, v5, v4, 0x7060302
-; GFX11-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v6
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v9, v11, vcc_lo
-; GFX11-NEXT: v_add3_u32 v9, v12, v3, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v13, v7, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v35
-; GFX11-NEXT: v_perm_b32 v8, v11, v6, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc_lo
-; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v12
-; GFX11-NEXT: v_add3_u32 v12, v13, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v35
-; GFX11-NEXT: v_bfe_u32 v15, v9, 16, 1
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 16, v3
-; GFX11-NEXT: v_dual_cndmask_b32 v7, v12, v13 :: v_dual_add_f32 v12, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_add3_u32 v13, v15, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v17, v12, 16, 1
-; GFX11-NEXT: v_perm_b32 v7, v7, v3, 0x7060302
-; GFX11-NEXT: v_lshrrev_b64 v[3:4], 24, v[0:1]
-; GFX11-NEXT: v_dual_cndmask_b32 v13, v13, v14 :: v_dual_add_f32 v14, 0x40c00000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v16, v17, v12, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v33
-; GFX11-NEXT: v_bfe_u32 v15, v10, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_bfe_u32 v19, v14, 16, 1
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v9, v15, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v10
-; GFX11-NEXT: v_add3_u32 v10, v19, v14, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v20, v9, v15 :: v_dual_add_f32 v9, 0x40c00000, v17
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v33
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v14
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v16, v18, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_bfe_u32 v16, v9, 16, 1
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v32
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v10, v15 :: v_dual_add_f32 v15, 0x40c00000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v18
-; GFX11-NEXT: v_add3_u32 v16, v16, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v17
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v12
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v16, v18, vcc_lo
-; GFX11-NEXT: v_add3_u32 v16, v19, v15, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v16, v19, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v32
-; GFX11-NEXT: v_perm_b32 v16, v19, v18, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v21, v22, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v19
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v17
-; GFX11-NEXT: v_bfe_u32 v9, v10, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v9, v9, v10, 0x7fff
-; GFX11-NEXT: v_perm_b32 v10, v20, v13, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v13, 8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v9, v23, vcc_lo
-; GFX11-NEXT: v_perm_b32 v9, v14, v12, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v11
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 24, v10
-; GFX11-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 8, v9
-; GFX11-NEXT: v_lshrrev_b64 v[27:28], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v15, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: .LBB54_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v38
-; GFX11-NEXT: v_mov_b32_e32 v4, v39
-; GFX11-NEXT: v_mov_b32_e32 v8, v36
-; GFX11-NEXT: v_mov_b32_e32 v12, v37
-; GFX11-NEXT: v_mov_b32_e32 v16, v34
-; GFX11-NEXT: v_mov_b32_e32 v20, v35
-; GFX11-NEXT: v_mov_b32_e32 v24, v32
-; GFX11-NEXT: v_mov_b32_e32 v28, v33
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <16 x bfloat> %a1 to <32 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x bfloat> %a to <32 x i8>
- br label %end
-
-end:
- %phi = phi <32 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i8> %phi
-}
-
-define <16 x bfloat> @bitcast_v32i8_to_v16bf16(<32 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i8_to_v16bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 24, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v17
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v25
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v3, v3, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v13, v13, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v18, v18, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v19
-; GCN-NEXT: v_or_b32_e32 v31, v1, v0
-; GCN-NEXT: v_or_b32_e32 v35, v36, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v3
-; GCN-NEXT: v_or_b32_e32 v3, v37, v4
-; GCN-NEXT: v_or_b32_e32 v33, v6, v5
-; GCN-NEXT: v_or_b32_e32 v5, v38, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v8
-; GCN-NEXT: v_or_b32_e32 v7, v39, v9
-; GCN-NEXT: v_or_b32_e32 v23, v11, v10
-; GCN-NEXT: v_or_b32_e32 v27, v48, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v13
-; GCN-NEXT: v_or_b32_e32 v11, v49, v14
-; GCN-NEXT: v_or_b32_e32 v32, v16, v15
-; GCN-NEXT: v_or_b32_e32 v13, v50, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v18
-; GCN-NEXT: v_or_b32_e32 v15, v51, v20
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: .LBB55_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB55_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v28
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v30
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v25
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v55, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v7, v11, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v54, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v19
-; GCN-NEXT: v_or_b32_e32 v16, v17, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v18
-; GCN-NEXT: v_or_b32_e32 v12, v53, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v10
-; GCN-NEXT: v_or_b32_e32 v4, v52, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x300, v3
-; GCN-NEXT: v_or_b32_e32 v3, v51, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v7
-; GCN-NEXT: v_or_b32_e32 v7, v50, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v13
-; GCN-NEXT: v_or_b32_e32 v11, v49, v15
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v15, v48, v17
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v14, v39, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v38, v9
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v6, v37, v6
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_or_b32_e32 v1, v36, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v7, v5
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v15, v13
-; GCN-NEXT: v_or_b32_e32 v10, v14, v12
-; GCN-NEXT: v_or_b32_e32 v8, v9, v8
-; GCN-NEXT: v_or_b32_e32 v4, v6, v4
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v10
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v6
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v2
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v1
-; GCN-NEXT: .LBB55_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v31
-; GCN-NEXT: v_mov_b32_e32 v1, v35
-; GCN-NEXT: v_mov_b32_e32 v2, v21
-; GCN-NEXT: v_mov_b32_e32 v4, v33
-; GCN-NEXT: v_mov_b32_e32 v6, v19
-; GCN-NEXT: v_mov_b32_e32 v8, v23
-; GCN-NEXT: v_mov_b32_e32 v9, v27
-; GCN-NEXT: v_mov_b32_e32 v10, v29
-; GCN-NEXT: v_mov_b32_e32 v12, v32
-; GCN-NEXT: v_mov_b32_e32 v14, v34
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i8_to_v16bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v33, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; VI-NEXT: v_mov_b32_e32 v32, v6
-; VI-NEXT: v_mov_b32_e32 v34, v4
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v9, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v2
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB55_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB55_4
-; VI-NEXT: .LBB55_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB55_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v32, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr8
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB55_2
-; VI-NEXT: .LBB55_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v30
-; VI-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v1, 0x300
-; VI-NEXT: v_add_u16_sdwa v7, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v28
-; VI-NEXT: v_or_b32_sdwa v28, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v26
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v24
-; VI-NEXT: v_or_b32_sdwa v24, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v22
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v20
-; VI-NEXT: v_or_b32_sdwa v20, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v16
-; VI-NEXT: v_or_b32_sdwa v16, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v14
-; VI-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v12
-; VI-NEXT: v_or_b32_sdwa v12, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v10
-; VI-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v34
-; VI-NEXT: v_or_b32_sdwa v10, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v33
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v2, v8, v2
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v12
-; VI-NEXT: v_add_u16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v31
-; VI-NEXT: v_or_b32_e32 v3, v8, v3
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v16
-; VI-NEXT: v_or_b32_sdwa v1, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_e32 v4, v8, v4
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v20
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v5, v8, v5
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v24
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v6, v8, v6
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v1, v1, v9
-; VI-NEXT: v_or_b32_e32 v7, v8, v7
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i8_to_v16bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v33, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v31, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v9, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v2
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB55_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB55_4
-; GFX9-NEXT: .LBB55_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB55_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v33, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_or_b32_sdwa v1, v31, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v1, v2, v1, s6
-; GFX9-NEXT: v_or_b32_sdwa v2, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr8
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr9
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB55_2
-; GFX9-NEXT: .LBB55_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v12
-; GFX9-NEXT: v_or_b32_sdwa v0, v13, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v0, v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v0, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v0, v37, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v35, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v10, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v9, v1, s6
-; GFX9-NEXT: v_perm_b32 v2, v8, v2, s6
-; GFX9-NEXT: v_perm_b32 v3, v11, v3, s6
-; GFX9-NEXT: v_perm_b32 v4, v15, v4, s6
-; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6
-; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6
-; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i8_to_v16bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v34, v2 :: v_dual_mov_b32 v33, v0
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:4
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: v_dual_mov_b32 v31, v6 :: v_dual_mov_b32 v32, v4
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v35, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v38, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v9, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v2
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB55_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB55_4
-; GFX11-NEXT: .LBB55_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB55_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v35
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v38
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v48
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v9
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v18
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GFX11-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v30
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v49
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v17
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v21
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v23
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v25
-; GFX11-NEXT: v_perm_b32 v3, v4, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v6, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v12, v11, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr8
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr9
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB55_2
-; GFX11-NEXT: .LBB55_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v30, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v20, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v23, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v25, v1
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v2, v21, v2
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v19, v3
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v4
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v17, v5
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v14, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v2
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v15, v0
-; GFX11-NEXT: v_add_nc_u16 v2, v8, 3
-; GFX11-NEXT: v_or_b32_e32 v1, v39, v1
-; GFX11-NEXT: v_add_nc_u16 v14, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v13, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v11, v4
-; GFX11-NEXT: v_add_nc_u16 v4, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v34, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v2, v48, v2
-; GFX11-NEXT: v_or_b32_e32 v4, v38, v4
-; GFX11-NEXT: v_or_b32_e32 v11, v37, v11
-; GFX11-NEXT: v_or_b32_e32 v13, v35, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v36, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_perm_b32 v0, v11, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v13, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v9, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v14, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v15, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v12, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v17, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i8> %a, splat (i8 3)
- %a2 = bitcast <32 x i8> %a1 to <16 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i8> %a to <16 x bfloat>
- br label %end
-
-end:
- %phi = phi <16 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x bfloat> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.288bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.288bit.ll
deleted file mode 100644
index 0adf547e19362..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.288bit.ll
+++ /dev/null
@@ -1,209 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <9 x float> @bitcast_v9i32_to_v9f32(<9 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v9i32_to_v9f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v9i32_to_v9f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v9i32_to_v9f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v9i32_to_v9f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v9
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <9 x i32> %a, splat (i32 3)
- %a2 = bitcast <9 x i32> %a1 to <9 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <9 x i32> %a to <9 x float>
- br label %end
-
-end:
- %phi = phi <9 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <9 x float> %phi
-}
-
-define <9 x i32> @bitcast_v9f32_to_v9i32(<9 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v9f32_to_v9i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v9f32_to_v9i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v9f32_to_v9i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v9f32_to_v9i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v9
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v7, 1.0, v7
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <9 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <9 x float> %a1 to <9 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <9 x float> %a to <9 x i32>
- br label %end
-
-end:
- %phi = phi <9 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <9 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
deleted file mode 100644
index f1e0c19f7fca3..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
+++ /dev/null
@@ -1,12523 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <10 x float> @bitcast_v10i32_to_v10f32(<10 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10i32_to_v10f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10i32_to_v10f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10i32_to_v10f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10i32_to_v10f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <10 x i32> %a, splat (i32 3)
- %a2 = bitcast <10 x i32> %a1 to <10 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x i32> %a to <10 x float>
- br label %end
-
-end:
- %phi = phi <10 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x float> %phi
-}
-
-define <10 x i32> @bitcast_v10f32_to_v10i32(<10 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10f32_to_v10i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10f32_to_v10i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10f32_to_v10i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10f32_to_v10i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <10 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <10 x float> %a1 to <10 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x float> %a to <10 x i32>
- br label %end
-
-end:
- %phi = phi <10 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x i32> %phi
-}
-
-define <20 x half> @bitcast_v10i32_to_v20f16(<10 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10i32_to_v20f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v29, v9
-; GCN-NEXT: v_mov_b32_e32 v28, v8
-; GCN-NEXT: v_mov_b32_e32 v27, v7
-; GCN-NEXT: v_mov_b32_e32 v26, v6
-; GCN-NEXT: v_mov_b32_e32 v25, v5
-; GCN-NEXT: v_mov_b32_e32 v24, v4
-; GCN-NEXT: v_mov_b32_e32 v23, v3
-; GCN-NEXT: v_mov_b32_e32 v22, v2
-; GCN-NEXT: v_mov_b32_e32 v21, v1
-; GCN-NEXT: v_mov_b32_e32 v20, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_4
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB2_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: .LBB2_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10i32_to_v20f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10i32_to_v20f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10i32_to_v20f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <10 x i32> %a, splat (i32 3)
- %a2 = bitcast <10 x i32> %a1 to <20 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x i32> %a to <20 x half>
- br label %end
-
-end:
- %phi = phi <20 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <20 x half> %phi
-}
-
-define <10 x i32> @bitcast_v20f16_to_v10i32(<20 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v20f16_to_v10i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v18
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_4
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB3_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v13
-; GCN-NEXT: v_or_b32_e32 v0, v29, v0
-; GCN-NEXT: v_or_b32_e32 v1, v27, v1
-; GCN-NEXT: v_or_b32_e32 v2, v25, v2
-; GCN-NEXT: v_or_b32_e32 v3, v23, v3
-; GCN-NEXT: v_or_b32_e32 v4, v22, v4
-; GCN-NEXT: v_or_b32_e32 v5, v21, v5
-; GCN-NEXT: v_or_b32_e32 v6, v20, v6
-; GCN-NEXT: v_or_b32_e32 v7, v12, v7
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: .LBB3_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v17, v16
-; GCN-NEXT: v_or_b32_e32 v6, v19, v18
-; GCN-NEXT: v_or_b32_e32 v7, v12, v15
-; GCN-NEXT: v_or_b32_e32 v8, v11, v14
-; GCN-NEXT: v_or_b32_e32 v9, v10, v13
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v20f16_to_v10i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v10, 0x200
-; VI-NEXT: v_add_f16_sdwa v11, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v11
-; VI-NEXT: v_add_f16_sdwa v11, v8, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v11
-; VI-NEXT: v_add_f16_sdwa v11, v7, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v11
-; VI-NEXT: v_add_f16_sdwa v11, v6, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v11
-; VI-NEXT: v_add_f16_sdwa v11, v5, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v11
-; VI-NEXT: v_add_f16_sdwa v11, v4, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v11
-; VI-NEXT: v_add_f16_sdwa v11, v3, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v11
-; VI-NEXT: v_add_f16_sdwa v11, v2, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_add_f16_sdwa v11, v1, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v10, v0, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v11
-; VI-NEXT: v_or_b32_e32 v0, v0, v10
-; VI-NEXT: .LBB3_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v20f16_to_v10i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB3_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB3_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v20f16_to_v10i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <20 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <20 x half> %a1 to <10 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <20 x half> %a to <10 x i32>
- br label %end
-
-end:
- %phi = phi <10 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x i32> %phi
-}
-
-define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10i32_to_v40i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v17, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB4_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v17, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB4_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v35
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v48
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v38
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v34
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v20
-; GCN-NEXT: v_or_b32_e32 v35, v49, v35
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 24, v28
-; GCN-NEXT: v_or_b32_e32 v48, v50, v48
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; GCN-NEXT: v_or_b32_e32 v3, v3, v29
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v22
-; GCN-NEXT: v_or_b32_e32 v4, v4, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_or_b32_e32 v5, v5, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v16
-; GCN-NEXT: v_or_b32_e32 v6, v6, v31
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v7, v7, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v24
-; GCN-NEXT: v_or_b32_e32 v9, v9, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v10, v10, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v35
-; GCN-NEXT: v_or_b32_e32 v19, v32, v33
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v25, v38, v39
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v27, v34, v36
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v21, v28, v30
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v15, v22, v17
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v12, v16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v18, v19
-; GCN-NEXT: v_or_b32_e32 v16, v24, v25
-; GCN-NEXT: v_or_b32_e32 v3, v3, v26
-; GCN-NEXT: v_or_b32_e32 v4, v4, v27
-; GCN-NEXT: v_or_b32_e32 v5, v5, v20
-; GCN-NEXT: v_or_b32_e32 v6, v6, v21
-; GCN-NEXT: v_or_b32_e32 v7, v7, v14
-; GCN-NEXT: v_or_b32_e32 v8, v8, v15
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v12
-; GCN-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10i32_to_v40i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB4_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB4_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; VI-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10i32_to_v40i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB4_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB4_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10i32_to_v40i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v11
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB4_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB4_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v16
-; GFX11-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v31, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v39
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v16
-; GFX11-NEXT: v_or_b32_e32 v15, v48, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v14, v35, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v31
-; GFX11-NEXT: v_or_b32_e32 v13, v30, v13
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_or_b32_e32 v16, v38, v37
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v13
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v29
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v27
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v26
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v32
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v34
-; GFX11-NEXT: v_or_b32_e32 v32, v33, v32
-; GFX11-NEXT: v_or_b32_e32 v12, v25, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v16
-; GFX11-NEXT: v_or_b32_e32 v11, v20, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v30
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v15
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <10 x i32> %a, splat (i32 3)
- %a2 = bitcast <10 x i32> %a1 to <40 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x i32> %a to <40 x i8>
- br label %end
-
-end:
- %phi = phi <40 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <40 x i8> %phi
-}
-
-define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v40i8_to_v10i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 24, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v40
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v55
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v14, v14, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v22, v22, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v26, v26, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v36, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v37, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v13, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v15, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v11, v17, v12
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v13, v19, v16
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v15, v21, v20
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v17, v23, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v19, v25, v27
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: v_or_b32_e32 v8, v16, v17
-; GCN-NEXT: v_or_b32_e32 v9, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: .LBB5_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v55
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v0, v38, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v49, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v50, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v51, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v14, v52, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v27, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v22, v29, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v26, v53, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v36, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v37, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v13, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v15, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_or_b32_e32 v11, v17, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v14
-; GCN-NEXT: v_or_b32_e32 v13, v19, v16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v18
-; GCN-NEXT: v_or_b32_e32 v15, v21, v20
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v22
-; GCN-NEXT: v_or_b32_e32 v17, v23, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 0x300, v26
-; GCN-NEXT: v_or_b32_e32 v19, v25, v27
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_or_b32_e32 v8, v17, v16
-; GCN-NEXT: v_or_b32_e32 v9, v19, v18
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x3000000, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v9
-; GCN-NEXT: .LBB5_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v40i8_to_v10i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v35, v8
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v4
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v8
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v44
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v39, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v38, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v37, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v36, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: .LBB5_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v9, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v33
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v35
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v53, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v52, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v51, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v50, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v49, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v48, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v7, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v25, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v7, v23, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: v_add_u16_e32 v7, 3, v28
-; VI-NEXT: v_add_u16_e32 v8, 3, v30
-; VI-NEXT: v_or_b32_sdwa v7, v21, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v19, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_add_u16_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v8, 3, v39
-; VI-NEXT: v_add_u16_e32 v10, 3, v38
-; VI-NEXT: v_or_b32_sdwa v8, v17, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v10, v15, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v10, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v8, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v37
-; VI-NEXT: v_add_u16_e32 v12, 3, v36
-; VI-NEXT: v_or_b32_sdwa v10, v13, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v9, v11, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: .LBB5_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v40i8_to_v10i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v35, v8
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v8
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v44
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v39, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v38, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v37, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v36, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: .LBB5_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v35
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v53, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v52, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v51, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v50, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v49, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v48, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v25, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v23, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v21, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v19, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v39
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v8, v17, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v9, v15, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v37
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v36
-; GFX9-NEXT: v_or_b32_sdwa v9, v13, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: .LBB5_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v40i8_to_v10i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v35, v8 :: v_dual_mov_b32 v34, v6
-; GFX11-NEXT: v_dual_mov_b32 v33, v4 :: v_dual_mov_b32 v32, v2
-; GFX11-NEXT: v_mov_b32_e32 v31, v0
-; GFX11-NEXT: s_clause 0x9
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v36, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v37, off, s32 offset:20
-; GFX11-NEXT: scratch_load_u16 v38, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(9)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(7)
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(5)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v66
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB5_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB5_4
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB5_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v53
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v54
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v64
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v65
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v48
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v49
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v50
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v51
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v52
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v23
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v25
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v27
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v29
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: .LBB5_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v18, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v53, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v54, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v55, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v64, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v65, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v48, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v49, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v50, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v51, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v52, v9
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_add_nc_u16 v5, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v12, v39, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v36, 3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v5, v21, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v23, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v25, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v29, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v19, v18
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <40 x i8> %a, splat (i8 3)
- %a2 = bitcast <40 x i8> %a1 to <10 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <40 x i8> %a to <10 x i32>
- br label %end
-
-end:
- %phi = phi <10 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x i32> %phi
-}
-
-define <5 x double> @bitcast_v10i32_to_v5f64(<10 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10i32_to_v5f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB6_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10i32_to_v5f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB6_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB6_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10i32_to_v5f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB6_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB6_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10i32_to_v5f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB6_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB6_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <10 x i32> %a, splat (i32 3)
- %a2 = bitcast <10 x i32> %a1 to <5 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x i32> %a to <5 x double>
- br label %end
-
-end:
- %phi = phi <5 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x double> %phi
-}
-
-define <10 x i32> @bitcast_v5f64_to_v10i32(<5 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f64_to_v10i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f64_to_v10i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f64_to_v10i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB7_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f64_to_v10i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB7_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <5 x double> %a1 to <10 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x double> %a to <10 x i32>
- br label %end
-
-end:
- %phi = phi <10 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x i32> %phi
-}
-
-define <5 x i64> @bitcast_v10i32_to_v5i64(<10 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10i32_to_v5i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10i32_to_v5i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB8_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB8_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10i32_to_v5i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB8_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB8_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10i32_to_v5i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB8_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB8_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <10 x i32> %a, splat (i32 3)
- %a2 = bitcast <10 x i32> %a1 to <5 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x i32> %a to <5 x i64>
- br label %end
-
-end:
- %phi = phi <5 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i64> %phi
-}
-
-define <10 x i32> @bitcast_v5i64_to_v10i32(<5 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i64_to_v10i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i64_to_v10i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i64_to_v10i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB9_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB9_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i64_to_v10i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i64> %a, splat (i64 3)
- %a2 = bitcast <5 x i64> %a1 to <10 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i64> %a to <10 x i32>
- br label %end
-
-end:
- %phi = phi <10 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x i32> %phi
-}
-
-define <20 x half> @bitcast_v10f32_to_v20f16(<10 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10f32_to_v20f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v29, v9
-; GCN-NEXT: v_mov_b32_e32 v28, v8
-; GCN-NEXT: v_mov_b32_e32 v27, v7
-; GCN-NEXT: v_mov_b32_e32 v26, v6
-; GCN-NEXT: v_mov_b32_e32 v25, v5
-; GCN-NEXT: v_mov_b32_e32 v24, v4
-; GCN-NEXT: v_mov_b32_e32 v23, v3
-; GCN-NEXT: v_mov_b32_e32 v22, v2
-; GCN-NEXT: v_mov_b32_e32 v21, v1
-; GCN-NEXT: v_mov_b32_e32 v20, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_4
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB10_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: .LBB10_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v21
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v23
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v25
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v27
-; GCN-NEXT: v_add_f32_e32 v17, 1.0, v28
-; GCN-NEXT: v_add_f32_e32 v19, 1.0, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10f32_to_v20f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB10_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB10_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10f32_to_v20f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB10_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB10_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10f32_to_v20f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <10 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <10 x float> %a1 to <20 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x float> %a to <20 x half>
- br label %end
-
-end:
- %phi = phi <20 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <20 x half> %phi
-}
-
-define <10 x float> @bitcast_v20f16_to_v10f32(<20 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v20f16_to_v10f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v18
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_4
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB11_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v13
-; GCN-NEXT: v_or_b32_e32 v0, v29, v0
-; GCN-NEXT: v_or_b32_e32 v1, v27, v1
-; GCN-NEXT: v_or_b32_e32 v2, v25, v2
-; GCN-NEXT: v_or_b32_e32 v3, v23, v3
-; GCN-NEXT: v_or_b32_e32 v4, v22, v4
-; GCN-NEXT: v_or_b32_e32 v5, v21, v5
-; GCN-NEXT: v_or_b32_e32 v6, v20, v6
-; GCN-NEXT: v_or_b32_e32 v7, v12, v7
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: v_or_b32_e32 v9, v10, v9
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: .LBB11_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v17, v16
-; GCN-NEXT: v_or_b32_e32 v6, v19, v18
-; GCN-NEXT: v_or_b32_e32 v7, v12, v15
-; GCN-NEXT: v_or_b32_e32 v8, v11, v14
-; GCN-NEXT: v_or_b32_e32 v9, v10, v13
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v20f16_to_v10f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v10, 0x200
-; VI-NEXT: v_add_f16_sdwa v11, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v11
-; VI-NEXT: v_add_f16_sdwa v11, v8, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v11
-; VI-NEXT: v_add_f16_sdwa v11, v7, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v11
-; VI-NEXT: v_add_f16_sdwa v11, v6, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v11
-; VI-NEXT: v_add_f16_sdwa v11, v5, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v11
-; VI-NEXT: v_add_f16_sdwa v11, v4, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v11
-; VI-NEXT: v_add_f16_sdwa v11, v3, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v11
-; VI-NEXT: v_add_f16_sdwa v11, v2, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_add_f16_sdwa v11, v1, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v10, v0, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v11
-; VI-NEXT: v_or_b32_e32 v0, v0, v10
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v20f16_to_v10f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v20f16_to_v10f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <20 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <20 x half> %a1 to <10 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <20 x half> %a to <10 x float>
- br label %end
-
-end:
- %phi = phi <10 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x float> %phi
-}
-
-define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10f32_to_v40i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v17, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB12_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v17, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB12_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v35
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v48
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v38
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v34
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v20
-; GCN-NEXT: v_or_b32_e32 v35, v49, v35
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 24, v28
-; GCN-NEXT: v_or_b32_e32 v48, v50, v48
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; GCN-NEXT: v_or_b32_e32 v3, v3, v29
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v22
-; GCN-NEXT: v_or_b32_e32 v4, v4, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_or_b32_e32 v5, v5, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 24, v16
-; GCN-NEXT: v_or_b32_e32 v6, v6, v31
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v7, v7, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v24
-; GCN-NEXT: v_or_b32_e32 v9, v9, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v10, v10, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v35
-; GCN-NEXT: v_or_b32_e32 v19, v32, v33
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v25, v38, v39
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v27, v34, v36
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v21, v28, v30
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v15, v22, v17
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v12, v16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v18, v19
-; GCN-NEXT: v_or_b32_e32 v16, v24, v25
-; GCN-NEXT: v_or_b32_e32 v3, v3, v26
-; GCN-NEXT: v_or_b32_e32 v4, v4, v27
-; GCN-NEXT: v_or_b32_e32 v5, v5, v20
-; GCN-NEXT: v_or_b32_e32 v6, v6, v21
-; GCN-NEXT: v_or_b32_e32 v7, v7, v14
-; GCN-NEXT: v_or_b32_e32 v8, v8, v15
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v12
-; GCN-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10f32_to_v40i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB12_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB12_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; VI-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10f32_to_v40i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB12_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB12_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10f32_to_v40i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v11
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB12_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v9, 1.0, v9
-; GFX11-NEXT: v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v7, 1.0, v7
-; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB12_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v16
-; GFX11-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v31, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v39
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v16
-; GFX11-NEXT: v_or_b32_e32 v15, v48, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v14, v35, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v31
-; GFX11-NEXT: v_or_b32_e32 v13, v30, v13
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_or_b32_e32 v16, v38, v37
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v13
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v29
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v27
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v26
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v32
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v34
-; GFX11-NEXT: v_or_b32_e32 v32, v33, v32
-; GFX11-NEXT: v_or_b32_e32 v12, v25, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v16
-; GFX11-NEXT: v_or_b32_e32 v11, v20, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v30
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v15
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <10 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <10 x float> %a1 to <40 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x float> %a to <40 x i8>
- br label %end
-
-end:
- %phi = phi <40 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <40 x i8> %phi
-}
-
-define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v40i8_to_v10f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:36
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 24, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v41
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v40
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v55
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v54
-; GCN-NEXT: v_or_b32_e32 v0, v0, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v10, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v14, v14, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v22, v22, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v26, v26, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v36, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v37, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v13, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v15, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v11, v17, v12
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v13, v19, v16
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v15, v21, v20
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v17, v23, v24
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v26
-; GCN-NEXT: v_or_b32_e32 v19, v25, v27
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: v_or_b32_e32 v8, v16, v17
-; GCN-NEXT: v_or_b32_e32 v9, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: .LBB13_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v55
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v54
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_or_b32_e32 v0, v38, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v49, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v50, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v51, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v14, v52, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v27, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v22, v29, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v26, v53, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v36, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v37, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v13, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v15, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_or_b32_e32 v11, v17, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v14
-; GCN-NEXT: v_or_b32_e32 v13, v19, v16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v18
-; GCN-NEXT: v_or_b32_e32 v15, v21, v20
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v22
-; GCN-NEXT: v_or_b32_e32 v17, v23, v24
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 0x300, v26
-; GCN-NEXT: v_or_b32_e32 v19, v25, v27
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_or_b32_e32 v8, v17, v16
-; GCN-NEXT: v_or_b32_e32 v9, v19, v18
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x3000000, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v9
-; GCN-NEXT: .LBB13_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v40i8_to_v10f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v35, v8
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v4
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_lshlrev_b16_e32 v11, 8, v8
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v13, 8, v44
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v10, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v12, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v14, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v39, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v38, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v37, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v36, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr10
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: .LBB13_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v9, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v33
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v35
-; VI-NEXT: v_add_u16_e32 v3, 3, v10
-; VI-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v12
-; VI-NEXT: v_add_u16_e32 v4, 3, v14
-; VI-NEXT: v_or_b32_sdwa v3, v53, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v52, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v51, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v50, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v49, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v48, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v7, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v25, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v7, v23, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: v_add_u16_e32 v7, 3, v28
-; VI-NEXT: v_add_u16_e32 v8, 3, v30
-; VI-NEXT: v_or_b32_sdwa v7, v21, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v19, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_add_u16_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v8, 3, v39
-; VI-NEXT: v_add_u16_e32 v10, 3, v38
-; VI-NEXT: v_or_b32_sdwa v8, v17, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v10, v15, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v10, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v8, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v37
-; VI-NEXT: v_add_u16_e32 v12, 3, v36
-; VI-NEXT: v_or_b32_sdwa v10, v13, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v9, v11, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: .LBB13_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v40i8_to_v10f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v35, v8
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshlrev_b16_e32 v11, 8, v8
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v13, 8, v44
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v10, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v12, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v14, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v39, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v38, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v37, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v36, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr10
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: .LBB13_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v35
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v10
-; GFX9-NEXT: v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v12
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v14
-; GFX9-NEXT: v_or_b32_sdwa v3, v53, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v52, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v51, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v50, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v49, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v48, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v25, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v23, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v21, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v19, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v39
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v8, v17, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v9, v15, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v37
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v36
-; GFX9-NEXT: v_or_b32_sdwa v9, v13, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: .LBB13_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v40i8_to_v10f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v35, v8 :: v_dual_mov_b32 v34, v6
-; GFX11-NEXT: v_dual_mov_b32 v33, v4 :: v_dual_mov_b32 v32, v2
-; GFX11-NEXT: v_mov_b32_e32 v31, v0
-; GFX11-NEXT: s_clause 0x9
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v36, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v37, off, s32 offset:20
-; GFX11-NEXT: scratch_load_u16 v38, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(9)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(7)
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(5)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v66
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_4
-; GFX11-NEXT: .LBB13_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB13_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v53
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v54
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v64
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v65
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v48
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v49
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v50
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v51
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v52
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v39
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v36
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v23
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v25
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v27
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v29
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr10
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB13_2
-; GFX11-NEXT: .LBB13_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v10, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v12, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v14, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v18, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v53, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v54, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v55, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v64, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v65, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v48, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v49, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v50, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v51, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v52, v9
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_add_nc_u16 v5, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v12, v39, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v36, 3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v5, v21, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v23, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v25, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v29, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v11, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v13, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v15, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v19, v18
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <40 x i8> %a, splat (i8 3)
- %a2 = bitcast <40 x i8> %a1 to <10 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <40 x i8> %a to <10 x float>
- br label %end
-
-end:
- %phi = phi <10 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x float> %phi
-}
-
-define <5 x double> @bitcast_v10f32_to_v5f64(<10 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10f32_to_v5f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10f32_to_v5f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB14_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB14_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10f32_to_v5f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB14_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB14_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10f32_to_v5f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <10 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <10 x float> %a1 to <5 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x float> %a to <5 x double>
- br label %end
-
-end:
- %phi = phi <5 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x double> %phi
-}
-
-define <10 x float> @bitcast_v5f64_to_v10f32(<5 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f64_to_v10f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f64_to_v10f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB15_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB15_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f64_to_v10f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB15_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB15_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f64_to_v10f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB15_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB15_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <5 x double> %a1 to <10 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x double> %a to <10 x float>
- br label %end
-
-end:
- %phi = phi <10 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x float> %phi
-}
-
-define <5 x i64> @bitcast_v10f32_to_v5i64(<10 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v10f32_to_v5i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v10f32_to_v5i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB16_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB16_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v10f32_to_v5i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB16_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB16_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v10f32_to_v5i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <10 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <10 x float> %a1 to <5 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <10 x float> %a to <5 x i64>
- br label %end
-
-end:
- %phi = phi <5 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i64> %phi
-}
-
-define <10 x float> @bitcast_v5i64_to_v10f32(<5 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i64_to_v10f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i64_to_v10f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i64_to_v10f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i64_to_v10f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i64> %a, splat (i64 3)
- %a2 = bitcast <5 x i64> %a1 to <10 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i64> %a to <10 x float>
- br label %end
-
-end:
- %phi = phi <10 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <10 x float> %phi
-}
-
-define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v20f16_to_v40i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v3
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v7
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v11
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v15
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v19
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GCN-NEXT: v_bfe_u32 v14, v21, 8, 8
-; GCN-NEXT: v_bfe_u32 v11, v4, 8, 8
-; GCN-NEXT: v_bfe_u32 v8, v3, 8, 8
-; GCN-NEXT: v_bfe_u32 v5, v2, 8, 8
-; GCN-NEXT: v_or_b32_e32 v29, v50, v6
-; GCN-NEXT: v_or_b32_e32 v27, v49, v7
-; GCN-NEXT: v_or_b32_e32 v20, v52, v9
-; GCN-NEXT: v_or_b32_e32 v17, v51, v10
-; GCN-NEXT: v_or_b32_e32 v13, v54, v12
-; GCN-NEXT: v_or_b32_e32 v12, v53, v15
-; GCN-NEXT: v_or_b32_e32 v10, v42, v16
-; GCN-NEXT: v_or_b32_e32 v9, v40, v18
-; GCN-NEXT: v_or_b32_e32 v7, v45, v19
-; GCN-NEXT: v_or_b32_e32 v6, v44, v22
-; GCN-NEXT: v_alignbit_b32 v34, v27, v29, 24
-; GCN-NEXT: v_alignbit_b32 v36, v27, v29, 16
-; GCN-NEXT: v_alignbit_b32 v38, v27, v29, 8
-; GCN-NEXT: v_alignbit_b32 v32, v17, v20, 24
-; GCN-NEXT: v_alignbit_b32 v33, v17, v20, 16
-; GCN-NEXT: v_alignbit_b32 v35, v17, v20, 8
-; GCN-NEXT: v_alignbit_b32 v26, v12, v13, 24
-; GCN-NEXT: v_alignbit_b32 v28, v12, v13, 16
-; GCN-NEXT: v_alignbit_b32 v31, v12, v13, 8
-; GCN-NEXT: v_alignbit_b32 v19, v9, v10, 24
-; GCN-NEXT: v_alignbit_b32 v24, v9, v10, 16
-; GCN-NEXT: v_alignbit_b32 v25, v9, v10, 8
-; GCN-NEXT: v_alignbit_b32 v15, v6, v7, 24
-; GCN-NEXT: v_alignbit_b32 v16, v6, v7, 16
-; GCN-NEXT: v_alignbit_b32 v18, v6, v7, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 8, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 8, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 8, v6
-; GCN-NEXT: v_bfe_u32 v22, v1, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: .LBB18_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v49
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v21
-; GCN-NEXT: v_bfe_u32 v14, v21, 8, 8
-; GCN-NEXT: v_bfe_u32 v11, v4, 8, 8
-; GCN-NEXT: v_bfe_u32 v8, v3, 8, 8
-; GCN-NEXT: v_bfe_u32 v5, v2, 8, 8
-; GCN-NEXT: v_or_b32_e32 v7, v6, v13
-; GCN-NEXT: v_or_b32_e32 v6, v22, v17
-; GCN-NEXT: v_or_b32_e32 v10, v9, v20
-; GCN-NEXT: v_or_b32_e32 v9, v23, v25
-; GCN-NEXT: v_or_b32_e32 v13, v12, v26
-; GCN-NEXT: v_or_b32_e32 v12, v24, v27
-; GCN-NEXT: v_or_b32_e32 v20, v15, v28
-; GCN-NEXT: v_or_b32_e32 v17, v16, v29
-; GCN-NEXT: v_or_b32_e32 v29, v18, v30
-; GCN-NEXT: v_or_b32_e32 v27, v19, v31
-; GCN-NEXT: v_alignbit_b32 v34, v27, v29, 24
-; GCN-NEXT: v_alignbit_b32 v36, v27, v29, 16
-; GCN-NEXT: v_alignbit_b32 v38, v27, v29, 8
-; GCN-NEXT: v_alignbit_b32 v32, v17, v20, 24
-; GCN-NEXT: v_alignbit_b32 v33, v17, v20, 16
-; GCN-NEXT: v_alignbit_b32 v35, v17, v20, 8
-; GCN-NEXT: v_alignbit_b32 v26, v12, v13, 24
-; GCN-NEXT: v_alignbit_b32 v28, v12, v13, 16
-; GCN-NEXT: v_alignbit_b32 v31, v12, v13, 8
-; GCN-NEXT: v_alignbit_b32 v19, v9, v10, 24
-; GCN-NEXT: v_alignbit_b32 v24, v9, v10, 16
-; GCN-NEXT: v_alignbit_b32 v25, v9, v10, 8
-; GCN-NEXT: v_alignbit_b32 v15, v6, v7, 24
-; GCN-NEXT: v_alignbit_b32 v16, v6, v7, 16
-; GCN-NEXT: v_alignbit_b32 v18, v6, v7, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 8, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 8, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 8, v6
-; GCN-NEXT: v_bfe_u32 v22, v1, 8, 8
-; GCN-NEXT: .LBB18_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v38
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v34
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v48
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v14
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v35
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v39
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_add_i32_e32 v52, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 24, v8
-; GCN-NEXT: v_add_i32_e32 v53, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v19
-; GCN-NEXT: v_add_i32_e32 v54, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 8, v30
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v5
-; GCN-NEXT: v_add_i32_e32 v55, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 24, v15
-; GCN-NEXT: v_add_i32_e32 v40, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v22
-; GCN-NEXT: v_add_i32_e32 v41, vcc, 36, v0
-; GCN-NEXT: v_or_b32_e32 v29, v29, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_or_b32_e32 v27, v27, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v35, v50, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_or_b32_e32 v17, v17, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_or_b32_e32 v13, v13, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_or_b32_e32 v12, v12, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v10, v10, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_or_b32_e32 v9, v9, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v7, v7, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v6, v6, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v23, v34, v36
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v21, v49, v21
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v35
-; GCN-NEXT: v_or_b32_e32 v29, v32, v33
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v4, v11, v4
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v13, v26, v28
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v3, v8, v3
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v10, v19, v24
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v15, v16
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v1, v22, v1
-; GCN-NEXT: v_or_b32_e32 v15, v18, v23
-; GCN-NEXT: v_or_b32_e32 v16, v25, v21
-; GCN-NEXT: v_or_b32_e32 v18, v27, v29
-; GCN-NEXT: v_or_b32_e32 v4, v17, v4
-; GCN-NEXT: v_or_b32_e32 v11, v11, v13
-; GCN-NEXT: v_or_b32_e32 v3, v12, v3
-; GCN-NEXT: v_or_b32_e32 v8, v8, v10
-; GCN-NEXT: v_or_b32_e32 v2, v9, v2
-; GCN-NEXT: v_or_b32_e32 v5, v5, v7
-; GCN-NEXT: v_or_b32_e32 v1, v6, v1
-; GCN-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v52, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v53, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v54, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v2, v55, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v40, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v1, v41, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v20f16_to_v40i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v1
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB18_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v29, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v27, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v30, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v35, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v38, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v48, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB18_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB18_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v11, 0x200
-; VI-NEXT: v_add_f16_sdwa v23, v2, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v23
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v26, v1, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v39, v2, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v26
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v21, v4, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v38, v1, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v21
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v25, v3, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v36, v4, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v25
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_add_f16_sdwa v19, v6, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v35, v3, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v19
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v24, v5, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v33, v6, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v24
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v18, v8, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v32, v5, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v18
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v22, v7, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v30, v8, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v22
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v17, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v20, v9, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v29, v7, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v17
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v20
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v49, v10, v12
-; VI-NEXT: v_or_b32_e32 v48, v9, v11
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[48:49]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[29:30]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[32:33]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[35:36]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[38:39]
-; VI-NEXT: v_lshrrev_b32_e32 v27, 8, v49
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v48
-; VI-NEXT: v_lshrrev_b32_e32 v30, 8, v30
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v29
-; VI-NEXT: v_lshrrev_b32_e32 v33, 8, v33
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v32
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v36
-; VI-NEXT: v_lshrrev_b32_e32 v37, 8, v35
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v39
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v38
-; VI-NEXT: v_bfe_u32 v29, v17, 8, 8
-; VI-NEXT: v_bfe_u32 v32, v18, 8, 8
-; VI-NEXT: v_bfe_u32 v35, v19, 8, 8
-; VI-NEXT: v_bfe_u32 v38, v21, 8, 8
-; VI-NEXT: v_bfe_u32 v48, v23, 8, 8
-; VI-NEXT: .LBB18_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; VI-NEXT: v_or_b32_sdwa v15, v26, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v48
-; VI-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v37
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; VI-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v38
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v21, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v24, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v33
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v35
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v19, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v22, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v30
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v28
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v29
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v17, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v20f16_to_v40i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB18_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB18_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB18_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB18_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v20f16_to_v40i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v11
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB18_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB18_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB18_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB18_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v16
-; GFX11-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v31, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v39
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v16
-; GFX11-NEXT: v_or_b32_e32 v15, v48, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v14, v35, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v31
-; GFX11-NEXT: v_or_b32_e32 v13, v30, v13
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_or_b32_e32 v16, v38, v37
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v13
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v29
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v27
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v26
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v32
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v34
-; GFX11-NEXT: v_or_b32_e32 v32, v33, v32
-; GFX11-NEXT: v_or_b32_e32 v12, v25, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v16
-; GFX11-NEXT: v_or_b32_e32 v11, v20, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v30
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v15
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <20 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <20 x half> %a1 to <40 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <20 x half> %a to <40 x i8>
- br label %end
-
-end:
- %phi = phi <40 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <40 x i8> %phi
-}
-
-define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v40i8_to_v20f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:36
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 8, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 8, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v31
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v56
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v57
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v58
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v59
-; GCN-NEXT: v_or_b32_e32 v0, v0, v36
-; GCN-NEXT: v_or_b32_e32 v1, v1, v37
-; GCN-NEXT: v_or_b32_e32 v2, v2, v38
-; GCN-NEXT: v_or_b32_e32 v3, v3, v39
-; GCN-NEXT: v_or_b32_e32 v4, v4, v48
-; GCN-NEXT: v_or_b32_e32 v5, v5, v49
-; GCN-NEXT: v_or_b32_e32 v6, v6, v50
-; GCN-NEXT: v_or_b32_e32 v7, v7, v51
-; GCN-NEXT: v_or_b32_e32 v8, v8, v52
-; GCN-NEXT: v_or_b32_e32 v9, v9, v53
-; GCN-NEXT: v_or_b32_e32 v10, v10, v54
-; GCN-NEXT: v_or_b32_e32 v11, v11, v55
-; GCN-NEXT: v_or_b32_e32 v12, v12, v40
-; GCN-NEXT: v_or_b32_e32 v13, v13, v41
-; GCN-NEXT: v_or_b32_e32 v14, v14, v42
-; GCN-NEXT: v_or_b32_e32 v15, v15, v43
-; GCN-NEXT: v_or_b32_e32 v16, v16, v44
-; GCN-NEXT: v_or_b32_e32 v17, v17, v45
-; GCN-NEXT: v_or_b32_e32 v18, v18, v46
-; GCN-NEXT: v_or_b32_e32 v19, v19, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: .LBB19_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v59
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v58
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v57
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v47, v1
-; GCN-NEXT: v_or_b32_e32 v3, v46, v3
-; GCN-NEXT: v_or_b32_e32 v5, v45, v5
-; GCN-NEXT: v_or_b32_e32 v7, v44, v7
-; GCN-NEXT: v_or_b32_e32 v9, v43, v9
-; GCN-NEXT: v_or_b32_e32 v11, v42, v11
-; GCN-NEXT: v_or_b32_e32 v13, v41, v13
-; GCN-NEXT: v_or_b32_e32 v15, v40, v15
-; GCN-NEXT: v_or_b32_e32 v17, v55, v17
-; GCN-NEXT: v_or_b32_e32 v19, v54, v19
-; GCN-NEXT: v_or_b32_e32 v18, v53, v18
-; GCN-NEXT: v_or_b32_e32 v16, v52, v16
-; GCN-NEXT: v_or_b32_e32 v14, v51, v14
-; GCN-NEXT: v_or_b32_e32 v12, v50, v12
-; GCN-NEXT: v_or_b32_e32 v10, v49, v10
-; GCN-NEXT: v_or_b32_e32 v8, v48, v8
-; GCN-NEXT: v_or_b32_e32 v6, v39, v6
-; GCN-NEXT: v_or_b32_e32 v4, v38, v4
-; GCN-NEXT: v_or_b32_e32 v2, v37, v2
-; GCN-NEXT: v_or_b32_e32 v0, v36, v0
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 0x300, v1
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v17
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v19
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v18
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v20
-; GCN-NEXT: .LBB19_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v31
-; GCN-NEXT: v_mov_b32_e32 v2, v23
-; GCN-NEXT: v_mov_b32_e32 v4, v35
-; GCN-NEXT: v_mov_b32_e32 v6, v27
-; GCN-NEXT: v_mov_b32_e32 v8, v33
-; GCN-NEXT: v_mov_b32_e32 v10, v21
-; GCN-NEXT: v_mov_b32_e32 v12, v25
-; GCN-NEXT: v_mov_b32_e32 v14, v29
-; GCN-NEXT: v_mov_b32_e32 v16, v32
-; GCN-NEXT: v_mov_b32_e32 v18, v34
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v40i8_to_v20f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v34, v10
-; VI-NEXT: v_mov_b32_e32 v33, v8
-; VI-NEXT: v_mov_b32_e32 v35, v6
-; VI-NEXT: v_mov_b32_e32 v38, v4
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v36, v0
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:4
-; VI-NEXT: v_mov_b32_e32 v31, v14
-; VI-NEXT: v_mov_b32_e32 v37, v12
-; VI-NEXT: v_lshlrev_b16_e32 v39, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v48, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v49, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v44, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v4
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v8
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v10
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v36, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v38, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v34, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v37, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v31, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v51, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v53, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v54, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v55, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: .LBB19_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v55
-; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v1, 0x300
-; VI-NEXT: v_add_u16_sdwa v9, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v0, 3, v54
-; VI-NEXT: v_or_b32_sdwa v10, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v0, 3, v53
-; VI-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v51
-; VI-NEXT: v_or_b32_sdwa v11, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v30
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v28
-; VI-NEXT: v_or_b32_sdwa v12, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v26
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v24
-; VI-NEXT: v_or_b32_sdwa v13, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v22
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v20
-; VI-NEXT: v_or_b32_sdwa v14, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v18
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v16
-; VI-NEXT: v_or_b32_sdwa v15, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v3, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v37
-; VI-NEXT: v_or_b32_sdwa v16, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v34
-; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v33
-; VI-NEXT: v_or_b32_sdwa v17, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v35
-; VI-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v38
-; VI-NEXT: v_or_b32_sdwa v19, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v36
-; VI-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v19
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v1, v1, v18
-; VI-NEXT: v_or_b32_e32 v2, v17, v2
-; VI-NEXT: v_or_b32_e32 v3, v16, v3
-; VI-NEXT: v_or_b32_e32 v4, v15, v4
-; VI-NEXT: v_or_b32_e32 v5, v14, v5
-; VI-NEXT: v_or_b32_e32 v6, v13, v6
-; VI-NEXT: v_or_b32_e32 v7, v12, v7
-; VI-NEXT: v_or_b32_e32 v8, v11, v8
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: .LBB19_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v40i8_to_v20f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v31, v10
-; GFX9-NEXT: v_mov_b32_e32 v32, v8
-; GFX9-NEXT: v_mov_b32_e32 v38, v6
-; GFX9-NEXT: v_mov_b32_e32 v35, v4
-; GFX9-NEXT: v_mov_b32_e32 v33, v2
-; GFX9-NEXT: v_mov_b32_e32 v36, v0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_mov_b32_e32 v37, v14
-; GFX9-NEXT: v_mov_b32_e32 v34, v12
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v39, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v44, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v47, 8, v10
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB19_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_or_b32_sdwa v1, v35, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v1, v2, v1, s6
-; GFX9-NEXT: v_or_b32_sdwa v2, v32, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v31, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: v_or_b32_sdwa v3, v34, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v37, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v55, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v53, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: v_or_b32_sdwa v9, v42, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v54, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: .LBB19_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB19_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v54
-; GFX9-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v55
-; GFX9-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v53
-; GFX9-NEXT: v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v37
-; GFX9-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v35
-; GFX9-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v36
-; GFX9-NEXT: v_add_u16_e32 v19, 3, v33
-; GFX9-NEXT: v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v19, v39, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v19
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v0, v19, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v18, v1, s6
-; GFX9-NEXT: v_perm_b32 v2, v17, v2, s6
-; GFX9-NEXT: v_perm_b32 v3, v16, v3, s6
-; GFX9-NEXT: v_perm_b32 v4, v15, v4, s6
-; GFX9-NEXT: v_perm_b32 v5, v14, v5, s6
-; GFX9-NEXT: v_perm_b32 v6, v13, v6, s6
-; GFX9-NEXT: v_perm_b32 v7, v12, v7, s6
-; GFX9-NEXT: v_perm_b32 v8, v11, v8, s6
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: .LBB19_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v40i8_to_v20f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v33, v10 :: v_dual_mov_b32 v34, v8
-; GFX11-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v35, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v2 :: v_dual_mov_b32 v36, v0
-; GFX11-NEXT: s_clause 0x9
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:20
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:4
-; GFX11-NEXT: v_dual_mov_b32 v31, v14 :: v_dual_mov_b32 v32, v12
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(9)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(7)
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(5)
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB19_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB19_4
-; GFX11-NEXT: .LBB19_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB19_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v48
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v49
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v51
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v50
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v52
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v54
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v17
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v66
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v25
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v69
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v27
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v29
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v70
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v71
-; GFX11-NEXT: v_perm_b32 v5, v6, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v10, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v12, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v14, v13, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB19_2
-; GFX11-NEXT: .LBB19_4: ; %cmp.true
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_add_nc_u16 v0, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v66, 3
-; GFX11-NEXT: s_waitcnt vmcnt(1)
-; GFX11-NEXT: v_add_nc_u16 v3, v65, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v28, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v2, v67, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v70, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v71, v1
-; GFX11-NEXT: v_or_b32_e32 v3, v29, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_or_b32_e32 v2, v27, v2
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v0, v69, v4
-; GFX11-NEXT: v_add_nc_u16 v1, v30, 3
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v3, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v26, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v2, v24, 3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v4, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v23, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_perm_b32 v8, v11, v8, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v25, v0
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v21, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v64, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v19, v4
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v31, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v54, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v53, v1
-; GFX11-NEXT: v_add_nc_u16 v2, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v17, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v52, v4
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v19, v33, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GFX11-NEXT: v_or_b32_e32 v2, v55, v2
-; GFX11-NEXT: v_or_b32_e32 v4, v51, v4
-; GFX11-NEXT: v_or_b32_e32 v16, v48, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v49, v17
-; GFX11-NEXT: v_or_b32_e32 v18, v50, v18
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v0
-; GFX11-NEXT: v_perm_b32 v0, v17, v16, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v18, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v19, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v20, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v21, v15, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v14, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v13, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v12, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v10, v9, 0x5040100
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <40 x i8> %a, splat (i8 3)
- %a2 = bitcast <40 x i8> %a1 to <20 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <40 x i8> %a to <20 x half>
- br label %end
-
-end:
- %phi = phi <20 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <20 x half> %phi
-}
-
-define <5 x double> @bitcast_v20f16_to_v5f64(<20 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v20f16_to_v5f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v18
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_4
-; GCN-NEXT: .LBB20_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB20_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v34
-; GCN-NEXT: v_or_b32_e32 v0, v35, v0
-; GCN-NEXT: v_or_b32_e32 v1, v33, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v19
-; GCN-NEXT: v_or_b32_e32 v2, v29, v2
-; GCN-NEXT: v_or_b32_e32 v3, v27, v3
-; GCN-NEXT: v_or_b32_e32 v4, v25, v4
-; GCN-NEXT: v_or_b32_e32 v5, v23, v5
-; GCN-NEXT: v_or_b32_e32 v6, v21, v6
-; GCN-NEXT: v_or_b32_e32 v7, v20, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v16, v9
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: .LBB20_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v16, v17
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v20f16_to_v5f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB20_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v10, 0x200
-; VI-NEXT: v_add_f16_sdwa v11, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v11
-; VI-NEXT: v_add_f16_sdwa v11, v8, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v11
-; VI-NEXT: v_add_f16_sdwa v11, v7, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v11
-; VI-NEXT: v_add_f16_sdwa v11, v6, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v11
-; VI-NEXT: v_add_f16_sdwa v11, v5, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v11
-; VI-NEXT: v_add_f16_sdwa v11, v4, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v11
-; VI-NEXT: v_add_f16_sdwa v11, v3, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v11
-; VI-NEXT: v_add_f16_sdwa v11, v2, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_add_f16_sdwa v11, v1, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v10, v0, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v11
-; VI-NEXT: v_or_b32_e32 v0, v0, v10
-; VI-NEXT: .LBB20_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v20f16_to_v5f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB20_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB20_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v20f16_to_v5f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB20_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB20_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <20 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <20 x half> %a1 to <5 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <20 x half> %a to <5 x double>
- br label %end
-
-end:
- %phi = phi <5 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x double> %phi
-}
-
-define <20 x half> @bitcast_v5f64_to_v20f16(<5 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f64_to_v20f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: .LBB21_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: .LBB21_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v24
-; GCN-NEXT: v_mov_b32_e32 v1, v29
-; GCN-NEXT: v_mov_b32_e32 v2, v20
-; GCN-NEXT: v_mov_b32_e32 v3, v28
-; GCN-NEXT: v_mov_b32_e32 v4, v21
-; GCN-NEXT: v_mov_b32_e32 v5, v27
-; GCN-NEXT: v_mov_b32_e32 v6, v22
-; GCN-NEXT: v_mov_b32_e32 v7, v26
-; GCN-NEXT: v_mov_b32_e32 v8, v23
-; GCN-NEXT: v_mov_b32_e32 v9, v25
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f64_to_v20f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f64_to_v20f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB21_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB21_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f64_to_v20f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB21_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB21_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <5 x double> %a1 to <20 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x double> %a to <20 x half>
- br label %end
-
-end:
- %phi = phi <20 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <20 x half> %phi
-}
-
-define <5 x i64> @bitcast_v20f16_to_v5i64(<20 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v20f16_to_v5i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v18
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_4
-; GCN-NEXT: .LBB22_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB22_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v34
-; GCN-NEXT: v_or_b32_e32 v0, v35, v0
-; GCN-NEXT: v_or_b32_e32 v1, v33, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v19
-; GCN-NEXT: v_or_b32_e32 v2, v29, v2
-; GCN-NEXT: v_or_b32_e32 v3, v27, v3
-; GCN-NEXT: v_or_b32_e32 v4, v25, v4
-; GCN-NEXT: v_or_b32_e32 v5, v23, v5
-; GCN-NEXT: v_or_b32_e32 v6, v21, v6
-; GCN-NEXT: v_or_b32_e32 v7, v20, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v16, v9
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: .LBB22_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v16, v17
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v20f16_to_v5i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB22_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v10, 0x200
-; VI-NEXT: v_add_f16_sdwa v11, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v11
-; VI-NEXT: v_add_f16_sdwa v11, v8, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v11
-; VI-NEXT: v_add_f16_sdwa v11, v7, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v11
-; VI-NEXT: v_add_f16_sdwa v11, v6, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v11
-; VI-NEXT: v_add_f16_sdwa v11, v5, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v11
-; VI-NEXT: v_add_f16_sdwa v11, v4, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v11
-; VI-NEXT: v_add_f16_sdwa v11, v3, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v11
-; VI-NEXT: v_add_f16_sdwa v11, v2, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v11
-; VI-NEXT: v_add_f16_sdwa v11, v1, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v10, v0, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v11
-; VI-NEXT: v_or_b32_e32 v0, v0, v10
-; VI-NEXT: .LBB22_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v20f16_to_v5i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB22_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB22_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v20f16_to_v5i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB22_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB22_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <20 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <20 x half> %a1 to <5 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <20 x half> %a to <5 x i64>
- br label %end
-
-end:
- %phi = phi <5 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i64> %phi
-}
-
-define <20 x half> @bitcast_v5i64_to_v20f16(<5 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i64_to_v20f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v22, v9
-; GCN-NEXT: v_mov_b32_e32 v21, v8
-; GCN-NEXT: v_mov_b32_e32 v24, v7
-; GCN-NEXT: v_mov_b32_e32 v23, v6
-; GCN-NEXT: v_mov_b32_e32 v26, v5
-; GCN-NEXT: v_mov_b32_e32 v25, v4
-; GCN-NEXT: v_mov_b32_e32 v28, v3
-; GCN-NEXT: v_mov_b32_e32 v27, v2
-; GCN-NEXT: v_mov_b32_e32 v29, v1
-; GCN-NEXT: v_mov_b32_e32 v20, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB23_4
-; GCN-NEXT: .LBB23_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB23_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v20
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: .LBB23_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v29, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v27
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v28, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v25
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v23
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v24, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v21
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v22, vcc
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v20
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i64_to_v20f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB23_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB23_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i64_to_v20f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB23_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB23_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i64_to_v20f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB23_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB23_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i64> %a, splat (i64 3)
- %a2 = bitcast <5 x i64> %a1 to <20 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i64> %a to <20 x half>
- br label %end
-
-end:
- %phi = phi <20 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <20 x half> %phi
-}
-
-define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v40i8_to_v5f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:36
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v47
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v46
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v45
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v44
-; GCN-NEXT: v_or_b32_e32 v0, v0, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v10, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v12, v12, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v16, v16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v49, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v50, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v17, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v11, v19, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v13, v21, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v15, v23, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v17, v25, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v27, v22
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: v_or_b32_e32 v8, v16, v17
-; GCN-NEXT: v_or_b32_e32 v9, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: .LBB24_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v51, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v53, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v54, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v55, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v40, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v12, v41, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v16, v42, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v43, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v49, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v50, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v17, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_or_b32_e32 v11, v19, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v13, v21, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_or_b32_e32 v15, v23, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v17, v25, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v20
-; GCN-NEXT: v_or_b32_e32 v19, v27, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_or_b32_e32 v8, v17, v16
-; GCN-NEXT: v_or_b32_e32 v9, v19, v18
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v9
-; GCN-NEXT: .LBB24_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v40i8_to_v5f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v36, v10
-; VI-NEXT: v_mov_b32_e32 v35, v8
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:4
-; VI-NEXT: v_mov_b32_e32 v38, v14
-; VI-NEXT: v_mov_b32_e32 v37, v12
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v44, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v4
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v8
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v10
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v36, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v37, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v38, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v50, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v49, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v39, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: .LBB24_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v9, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v33
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v35
-; VI-NEXT: v_add_u16_e32 v3, 3, v36
-; VI-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v43, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v37
-; VI-NEXT: v_add_u16_e32 v4, 3, v38
-; VI-NEXT: v_or_b32_sdwa v3, v42, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v41, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v40, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v55, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v7, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v52, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v7, v51, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: v_add_u16_e32 v7, 3, v28
-; VI-NEXT: v_add_u16_e32 v8, 3, v30
-; VI-NEXT: v_or_b32_sdwa v7, v27, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v25, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_add_u16_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v8, 3, v50
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v8, v23, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v10, v21, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v10, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v8, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v48
-; VI-NEXT: v_add_u16_e32 v11, 3, v39
-; VI-NEXT: v_or_b32_sdwa v10, v19, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v11, v17, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v9, v11, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: .LBB24_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v40i8_to_v5f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v36, v10
-; GFX9-NEXT: v_mov_b32_e32 v35, v8
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_mov_b32_e32 v38, v14
-; GFX9-NEXT: v_mov_b32_e32 v37, v12
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v47, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v44, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v10
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v38, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v50, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v49, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v39, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: .LBB24_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v35
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v36
-; GFX9-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v43, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v37
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v3, v42, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v41, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v40, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v55, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v52, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v51, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v27, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v25, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v50
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v8, v23, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v9, v21, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v48
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v9, v19, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v17, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: .LBB24_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v40i8_to_v5f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x9
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:20
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:4
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v68, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(9)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(7)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(5)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB24_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB24_4
-; GFX11-NEXT: .LBB24_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB24_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v67
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v68
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v69
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v70
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v71
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v54
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v65
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v66
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v51
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v52
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v53
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v27
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v29
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v17
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v23
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_2
-; GFX11-NEXT: .LBB24_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v18, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v67, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v68, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v69, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v70, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v71, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v54, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v55, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v64, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v65, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v66, v9
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_add_nc_u16 v5, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v11, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v51, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v52, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v53, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v29, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v17, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v19, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v21, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v23, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v25, v14
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <40 x i8> %a, splat (i8 3)
- %a2 = bitcast <40 x i8> %a1 to <5 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <40 x i8> %a to <5 x double>
- br label %end
-
-end:
- %phi = phi <5 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x double> %phi
-}
-
-define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f64_to_v40i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v16, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v17, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v18, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v19, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v30, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB25_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v16, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v17, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v18, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v19, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v30, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v31, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB25_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v35
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v48
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v38
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v34
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_add_i32_e32 v38, vcc, 12, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v17
-; GCN-NEXT: v_or_b32_e32 v49, v49, v50
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 8, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_or_b32_e32 v48, v51, v48
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 20, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; GCN-NEXT: v_or_b32_e32 v3, v3, v31
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 24, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v23
-; GCN-NEXT: v_or_b32_e32 v4, v4, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 28, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_or_b32_e32 v5, v5, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 32, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 8, v22
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v20
-; GCN-NEXT: v_or_b32_e32 v6, v6, v28
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 36, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_or_b32_e32 v7, v7, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v24
-; GCN-NEXT: v_or_b32_e32 v9, v9, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v10, v10, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v21
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v49
-; GCN-NEXT: v_or_b32_e32 v22, v32, v35
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v25, v33, v39
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v29, v29, v30
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v30, v34, v36
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v17, v17, v18
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v18, v26, v27
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v15, v23, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v12, v20, v13
-; GCN-NEXT: v_or_b32_e32 v13, v21, v22
-; GCN-NEXT: v_or_b32_e32 v16, v24, v25
-; GCN-NEXT: v_or_b32_e32 v3, v3, v29
-; GCN-NEXT: v_or_b32_e32 v4, v4, v30
-; GCN-NEXT: v_or_b32_e32 v5, v5, v17
-; GCN-NEXT: v_or_b32_e32 v6, v6, v18
-; GCN-NEXT: v_or_b32_e32 v7, v7, v14
-; GCN-NEXT: v_or_b32_e32 v8, v8, v15
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v12
-; GCN-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v38, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v31, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v28, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f64_to_v40i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB25_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; VI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; VI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; VI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; VI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB25_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; VI-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f64_to_v40i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB25_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX9-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX9-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX9-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX9-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB25_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f64_to_v40i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v11
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB25_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB25_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v16
-; GFX11-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v31, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v39
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v16
-; GFX11-NEXT: v_or_b32_e32 v15, v48, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v14, v35, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v31
-; GFX11-NEXT: v_or_b32_e32 v13, v30, v13
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_or_b32_e32 v16, v38, v37
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v13
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v29
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v27
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v26
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v32
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v34
-; GFX11-NEXT: v_or_b32_e32 v32, v33, v32
-; GFX11-NEXT: v_or_b32_e32 v12, v25, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v16
-; GFX11-NEXT: v_or_b32_e32 v11, v20, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v30
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v15
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <5 x double> %a1 to <40 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x double> %a to <40 x i8>
- br label %end
-
-end:
- %phi = phi <40 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <40 x i8> %phi
-}
-
-define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v40i8_to_v5i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:36
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 8, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v47
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v46
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v45
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v44
-; GCN-NEXT: v_or_b32_e32 v0, v0, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v2, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v4, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v6, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v8, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v10, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v12, v12, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v16, v16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v20, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v5, v49, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v7, v50, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v9, v17, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v11, v19, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v13, v21, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v15, v23, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v17, v25, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GCN-NEXT: v_or_b32_e32 v19, v27, v22
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v1, v2, v3
-; GCN-NEXT: v_or_b32_e32 v2, v4, v5
-; GCN-NEXT: v_or_b32_e32 v3, v6, v7
-; GCN-NEXT: v_or_b32_e32 v4, v8, v9
-; GCN-NEXT: v_or_b32_e32 v5, v10, v11
-; GCN-NEXT: v_or_b32_e32 v6, v12, v13
-; GCN-NEXT: v_or_b32_e32 v7, v14, v15
-; GCN-NEXT: v_or_b32_e32 v8, v16, v17
-; GCN-NEXT: v_or_b32_e32 v9, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: .LBB26_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_movk_i32 s6, 0x300
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: s_mov_b32 s7, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v44
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_or_b32_e32 v0, v51, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_or_b32_e32 v4, v53, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_or_b32_e32 v6, v54, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_or_b32_e32 v8, v55, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_or_b32_e32 v10, v40, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_or_b32_e32 v12, v41, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v16, v42, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_or_b32_e32 v20, v43, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_or_b32_e32 v1, v39, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v49, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_or_b32_e32 v7, v50, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_or_b32_e32 v9, v17, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_or_b32_e32 v11, v19, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_or_b32_e32 v13, v21, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; GCN-NEXT: v_or_b32_e32 v15, v23, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; GCN-NEXT: v_or_b32_e32 v17, v25, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v20
-; GCN-NEXT: v_or_b32_e32 v19, v27, v22
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_or_b32_e32 v2, v5, v4
-; GCN-NEXT: v_or_b32_e32 v3, v7, v6
-; GCN-NEXT: v_or_b32_e32 v4, v9, v8
-; GCN-NEXT: v_or_b32_e32 v5, v11, v10
-; GCN-NEXT: v_or_b32_e32 v6, v13, v12
-; GCN-NEXT: v_or_b32_e32 v7, v15, v14
-; GCN-NEXT: v_or_b32_e32 v8, v17, v16
-; GCN-NEXT: v_or_b32_e32 v9, v19, v18
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v9
-; GCN-NEXT: .LBB26_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v40i8_to_v5i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v36, v10
-; VI-NEXT: v_mov_b32_e32 v35, v8
-; VI-NEXT: v_mov_b32_e32 v34, v6
-; VI-NEXT: v_mov_b32_e32 v33, v4
-; VI-NEXT: v_mov_b32_e32 v32, v2
-; VI-NEXT: v_mov_b32_e32 v31, v0
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:4
-; VI-NEXT: v_mov_b32_e32 v38, v14
-; VI-NEXT: v_mov_b32_e32 v37, v12
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v44, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v2
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v4
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v8
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v10
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB26_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v31, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v32, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v33, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v34, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v36, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v37, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v38, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v16, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v18, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v20, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v22, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v24, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v26, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v30, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v50, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v49, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v39, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: .LBB26_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB26_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v31
-; VI-NEXT: v_add_u16_e32 v1, 3, v32
-; VI-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v9, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_add_u16_e32 v1, 3, v33
-; VI-NEXT: v_add_u16_e32 v2, 3, v34
-; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 0x300, v1
-; VI-NEXT: v_add_u16_sdwa v2, v2, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v1, v2
-; VI-NEXT: v_add_u16_e32 v2, 3, v35
-; VI-NEXT: v_add_u16_e32 v3, 3, v36
-; VI-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v43, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: v_add_u16_e32 v3, 3, v37
-; VI-NEXT: v_add_u16_e32 v4, 3, v38
-; VI-NEXT: v_or_b32_sdwa v3, v42, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v4, v41, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_add_u16_sdwa v4, v4, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: v_add_u16_e32 v4, 3, v16
-; VI-NEXT: v_add_u16_e32 v5, 3, v18
-; VI-NEXT: v_or_b32_sdwa v4, v40, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v5, v55, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_add_u16_sdwa v5, v5, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: v_add_u16_e32 v5, 3, v20
-; VI-NEXT: v_add_u16_e32 v6, 3, v22
-; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_add_u16_sdwa v6, v6, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: v_add_u16_e32 v6, 3, v24
-; VI-NEXT: v_add_u16_e32 v7, 3, v26
-; VI-NEXT: v_or_b32_sdwa v6, v52, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v7, v51, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_add_u16_sdwa v7, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: v_add_u16_e32 v7, 3, v28
-; VI-NEXT: v_add_u16_e32 v8, 3, v30
-; VI-NEXT: v_or_b32_sdwa v7, v27, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v8, v25, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_add_u16_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v8, 3, v50
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v8, v23, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v10, v21, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_add_u16_sdwa v10, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v8, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v48
-; VI-NEXT: v_add_u16_e32 v11, 3, v39
-; VI-NEXT: v_or_b32_sdwa v10, v19, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v11, v17, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_add_u16_sdwa v9, v11, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v10, v9
-; VI-NEXT: .LBB26_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v40i8_to_v5i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v36, v10
-; GFX9-NEXT: v_mov_b32_e32 v35, v8
-; GFX9-NEXT: v_mov_b32_e32 v34, v6
-; GFX9-NEXT: v_mov_b32_e32 v33, v4
-; GFX9-NEXT: v_mov_b32_e32 v32, v2
-; GFX9-NEXT: v_mov_b32_e32 v31, v0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_mov_b32_e32 v38, v14
-; GFX9-NEXT: v_mov_b32_e32 v37, v12
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v47, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v44, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v42, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v52, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v51, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v29
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v10
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB26_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v34, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v38, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v18, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v22, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v24, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v26, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v30, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v50, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v49, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v39, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: .LBB26_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB26_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v31
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v33
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v1
-; GFX9-NEXT: v_add_u16_sdwa v2, v2, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v35
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v36
-; GFX9-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v43, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v37
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v3, v42, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v4, v41, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v16
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v4, v40, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v5, v55, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v20
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v24
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v6, v52, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v7, v51, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v28
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v7, v27, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v8, v25, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v50
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v8, v23, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v9, v21, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v48
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v9, v19, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v10, v17, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: .LBB26_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v40i8_to_v5i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x9
-; GFX11-NEXT: scratch_load_b32 v0, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:20
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:4
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v68, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v70, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v64, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(9)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(7)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(5)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB26_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB26_4
-; GFX11-NEXT: .LBB26_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB26_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v18
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v67
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v68
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v69
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v70
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v71
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v54
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v55
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v64
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v65
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v66
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v51
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v52
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v53
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v27
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v29
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v17
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v23
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB26_2
-; GFX11-NEXT: .LBB26_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v18, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v67, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v68, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v69, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v70, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v71, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v54, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v55, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v64, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v65, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v66, v9
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v4, v8, v9
-; GFX11-NEXT: v_add_nc_u16 v5, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v22, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v11, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v51, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v52, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v53, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v27, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v29, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v17, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v19, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v21, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v23, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v25, v14
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <40 x i8> %a, splat (i8 3)
- %a2 = bitcast <40 x i8> %a1 to <5 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <40 x i8> %a to <5 x i64>
- br label %end
-
-end:
- %phi = phi <5 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i64> %phi
-}
-
-define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i64_to_v40i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v16, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB27_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_alignbit_b32 v11, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v12, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v14, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v15, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v16, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v20, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v21, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v23, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v26, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v27, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v29, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v32, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v33, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v35, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 8, v2
-; GCN-NEXT: .LBB27_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v49, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 8, v35
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 8, v48
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v38
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v34
-; GCN-NEXT: v_add_i32_e32 v51, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v20
-; GCN-NEXT: v_or_b32_e32 v35, v49, v35
-; GCN-NEXT: v_add_i32_e32 v49, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 24, v28
-; GCN-NEXT: v_or_b32_e32 v48, v50, v48
-; GCN-NEXT: v_add_i32_e32 v50, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; GCN-NEXT: v_or_b32_e32 v3, v3, v29
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v22
-; GCN-NEXT: v_or_b32_e32 v4, v4, v37
-; GCN-NEXT: v_add_i32_e32 v37, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; GCN-NEXT: v_or_b32_e32 v5, v5, v23
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 24, v17
-; GCN-NEXT: v_or_b32_e32 v6, v6, v31
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 36, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: v_or_b32_e32 v7, v7, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_or_b32_e32 v8, v8, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v24
-; GCN-NEXT: v_or_b32_e32 v9, v9, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_or_b32_e32 v10, v10, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v18
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v35
-; GCN-NEXT: v_or_b32_e32 v19, v32, v33
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v25, v38, v39
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v27, v34, v36
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_or_b32_e32 v21, v28, v30
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v14, v14, v15
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_or_b32_e32 v15, v22, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_or_b32_e32 v12, v17, v13
-; GCN-NEXT: v_or_b32_e32 v13, v18, v19
-; GCN-NEXT: v_or_b32_e32 v16, v24, v25
-; GCN-NEXT: v_or_b32_e32 v3, v3, v26
-; GCN-NEXT: v_or_b32_e32 v4, v4, v27
-; GCN-NEXT: v_or_b32_e32 v5, v5, v20
-; GCN-NEXT: v_or_b32_e32 v6, v6, v21
-; GCN-NEXT: v_or_b32_e32 v7, v7, v14
-; GCN-NEXT: v_or_b32_e32 v8, v8, v15
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v12
-; GCN-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v4, v51, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v5, v49, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v6, v50, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v29, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v37, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v23, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v31, s[0:3], 0 offen
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i64_to_v40i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr14
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr12
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB27_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; VI-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; VI-NEXT: .LBB27_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; VI-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i64_to_v40i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr14
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr13
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr12
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB27_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v1, vcc, 3, v1
-; GFX9-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, 3, v3
-; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, 3, v5
-; GFX9-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v7, vcc, 3, v7
-; GFX9-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v8, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v9, vcc, 3, v9
-; GFX9-NEXT: v_addc_co_u32_e32 v10, vcc, 0, v10, vcc
-; GFX9-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: .LBB27_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v15, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v14
-; GFX9-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v27
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v26
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v12
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v25, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v24
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v23, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v20, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i64_to_v40i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v11
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr15
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr14
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr13
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr12
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr11
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB27_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, v3, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, v5, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v7, vcc_lo, v7, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v9, vcc_lo, v9, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v1, vcc_lo, v1, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo
-; GFX11-NEXT: v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[12:13], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v21, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v22, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v24, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v16, 8, v1
-; GFX11-NEXT: .LBB27_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v16
-; GFX11-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v36, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-NEXT: v_lshlrev_b16 v14, 8, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v31, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v39
-; GFX11-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-NEXT: v_lshlrev_b16 v37, 8, v37
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v16
-; GFX11-NEXT: v_or_b32_e32 v15, v48, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v36
-; GFX11-NEXT: v_or_b32_e32 v14, v35, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v31
-; GFX11-NEXT: v_or_b32_e32 v13, v30, v13
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-NEXT: v_or_b32_e32 v16, v38, v37
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v15
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v14
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v13
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v13, 8, v29
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v28
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v27
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v16
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v26
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_lshlrev_b16 v34, 8, v34
-; GFX11-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-NEXT: v_lshlrev_b16 v32, 8, v32
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-NEXT: v_lshlrev_b16 v12, 8, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v23
-; GFX11-NEXT: v_lshlrev_b16 v15, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v16
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v16, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_lshlrev_b16 v11, 8, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v34
-; GFX11-NEXT: v_or_b32_e32 v32, v33, v32
-; GFX11-NEXT: v_or_b32_e32 v12, v25, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v16
-; GFX11-NEXT: v_or_b32_e32 v11, v20, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v19
-; GFX11-NEXT: v_or_b32_e32 v15, v18, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v30
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v13
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v12
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v14
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v11
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v15
-; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i64> %a, splat (i64 3)
- %a2 = bitcast <5 x i64> %a1 to <40 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i64> %a to <40 x i8>
- br label %end
-
-end:
- %phi = phi <40 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <40 x i8> %phi
-}
-
-define <5 x i64> @bitcast_v5f64_to_v5i64(<5 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5f64_to_v5i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: .LBB28_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5f64_to_v5i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB28_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: .LBB28_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5f64_to_v5i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB28_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: .LBB28_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5f64_to_v5i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: .LBB28_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <5 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <5 x double> %a1 to <5 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x double> %a to <5 x i64>
- br label %end
-
-end:
- %phi = phi <5 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x i64> %phi
-}
-
-define <5 x double> @bitcast_v5i64_to_v5f64(<5 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v5i64_to_v5f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: .LBB29_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v5i64_to_v5f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: .LBB29_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v5i64_to_v5f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB29_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: .LBB29_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v5i64_to_v5f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v10
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: .LBB29_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <5 x i64> %a, splat (i64 3)
- %a2 = bitcast <5 x i64> %a1 to <5 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <5 x i64> %a to <5 x double>
- br label %end
-
-end:
- %phi = phi <5 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <5 x double> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
deleted file mode 100644
index 1c51395128917..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
+++ /dev/null
@@ -1,4926 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define float @bitcast_i32_to_f32(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define i32 @bitcast_f32_to_i32(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <2 x i16> @bitcast_i32_to_v2i16(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_4
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB2_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: .LBB2_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define i32 @bitcast_v2i16_to_i32(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v3, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_4
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB3_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: .LBB3_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 3
-; VI-NEXT: v_add_u16_e32 v1, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <2 x half> @bitcast_i32_to_v2f16(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB4_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB4_4
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB4_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: .LBB4_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define i32 @bitcast_v2f16_to_i32(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_4
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB5_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: .LBB5_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 0x200
-; VI-NEXT: v_add_f16_sdwa v1, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <2 x bfloat> @bitcast_i32_to_v2bf16(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB6_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB6_4
-; GCN-NEXT: .LBB6_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB6_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: .LBB6_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define i32 @bitcast_v2bf16_to_i32(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_4
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB7_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: .LBB7_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v1, 16
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v1, s6
-; GFX9-NEXT: .LBB7_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
-; GFX11-NEXT: .LBB7_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <1 x i32> @bitcast_i32_to_v1i32(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
-
-define i32 @bitcast_v1i32_to_i32(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <4 x i8> @bitcast_i32_to_v4i8(i32 %a, i32 %b) {
-; GCN-LABEL: bitcast_i32_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_4
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB10_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: .LBB10_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_i32_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB10_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB10_4
-; VI-NEXT: .LBB10_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB10_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB10_2
-; VI-NEXT: .LBB10_4: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_i32_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB10_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB10_4
-; GFX9-NEXT: .LBB10_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB10_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB10_2
-; GFX9-NEXT: .LBB10_4: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_i32_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB10_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB10_4
-; GFX11-NEXT: .LBB10_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB10_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB10_2
-; GFX11-NEXT: .LBB10_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add i32 %a, 3
- %a2 = bitcast i32 %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast i32 %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB11_4
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB11_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: .LBB11_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x3000000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB11_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB11_4
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB11_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: .LBB11_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB11_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB11_4
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB11_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: .LBB11_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB11_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB11_4
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB11_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: .LBB11_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to i32
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to i32
- br label %end
-
-end:
- %phi = phi i32 [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret i32 %phi
-}
-
-define <2 x i16> @bitcast_f32_to_v2i16(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB12_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB12_4
-; GCN-NEXT: .LBB12_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB12_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: .LBB12_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define float @bitcast_v2i16_to_f32(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v3, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB13_4
-; GCN-NEXT: .LBB13_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB13_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: .LBB13_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 3
-; VI-NEXT: v_add_u16_e32 v1, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define <2 x half> @bitcast_f32_to_v2f16(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB14_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB14_4
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB14_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: .LBB14_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define float @bitcast_v2f16_to_f32(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB15_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB15_4
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB15_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: .LBB15_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 0x200
-; VI-NEXT: v_add_f16_sdwa v1, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define <2 x bfloat> @bitcast_f32_to_v2bf16(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB16_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB16_4
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB16_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: .LBB16_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define float @bitcast_v2bf16_to_f32(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB17_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB17_4
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB17_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: .LBB17_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v1, 16
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v1, s6
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define <1 x i32> @bitcast_f32_to_v1i32(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB18_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
-
-define float @bitcast_v1i32_to_f32(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB19_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define <4 x i8> @bitcast_f32_to_v4i8(float %a, i32 %b) {
-; GCN-LABEL: bitcast_f32_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB20_4
-; GCN-NEXT: .LBB20_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB20_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: .LBB20_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_f32_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB20_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB20_4
-; VI-NEXT: .LBB20_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB20_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB20_2
-; VI-NEXT: .LBB20_4: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_f32_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB20_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB20_4
-; GFX9-NEXT: .LBB20_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB20_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB20_2
-; GFX9-NEXT: .LBB20_4: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_f32_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB20_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB20_4
-; GFX11-NEXT: .LBB20_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB20_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB20_2
-; GFX11-NEXT: .LBB20_4: ; %cmp.true
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd float %a, 1.000000e+00
- %a2 = bitcast float %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast float %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB21_4
-; GCN-NEXT: .LBB21_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB21_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: .LBB21_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x3000000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB21_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB21_4
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB21_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: .LBB21_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB21_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB21_4
-; GFX9-NEXT: .LBB21_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB21_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB21_2
-; GFX9-NEXT: .LBB21_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB21_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB21_4
-; GFX11-NEXT: .LBB21_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB21_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB21_2
-; GFX11-NEXT: .LBB21_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to float
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to float
- br label %end
-
-end:
- %phi = phi float [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret float %phi
-}
-
-define <2 x half> @bitcast_v2i16_to_v2f16(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_mov_b32_e32 v3, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_4
-; GCN-NEXT: .LBB22_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB22_3: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: .LBB22_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 3
-; VI-NEXT: v_add_u16_sdwa v1, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define <2 x i16> @bitcast_v2f16_to_v2i16(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: .LBB23_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 0x200
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define <2 x bfloat> @bitcast_v2i16_to_v2bf16(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; GCN-NEXT: .LBB24_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 3
-; VI-NEXT: v_add_u16_sdwa v1, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define <2 x i16> @bitcast_v2bf16_to_v2i16(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB25_4
-; GCN-NEXT: .LBB25_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB25_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: .LBB25_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v1, 16
-; VI-NEXT: .LBB25_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v1, s6
-; GFX9-NEXT: .LBB25_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
-; GFX11-NEXT: .LBB25_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define <1 x i32> @bitcast_v2i16_to_v1i32(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v3, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB26_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB26_4
-; GCN-NEXT: .LBB26_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB26_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: .LBB26_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 3
-; VI-NEXT: v_add_u16_e32 v1, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
-
-define <2 x i16> @bitcast_v1i32_to_v2i16(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB27_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB27_4
-; GCN-NEXT: .LBB27_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB27_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: .LBB27_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define <4 x i8> @bitcast_v2i16_to_v4i8(<2 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2i16_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB28_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB28_4
-; GCN-NEXT: .LBB28_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB28_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v4
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: v_bfe_u32 v3, v4, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: .LBB28_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v4
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: v_bfe_u32 v3, v3, 8, 8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2i16_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v4, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v4
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB28_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB28_4
-; VI-NEXT: .LBB28_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB28_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v4
-; VI-NEXT: v_mov_b32_e32 v0, v4
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB28_2
-; VI-NEXT: .LBB28_4: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 3
-; VI-NEXT: v_add_u16_sdwa v2, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v4
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; VI-NEXT: v_or_b32_e32 v1, v0, v1
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; VI-NEXT: v_bfe_u32 v3, v2, 8, 8
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2i16_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB28_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB28_4
-; GFX9-NEXT: .LBB28_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB28_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB28_2
-; GFX9-NEXT: .LBB28_4: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2i16_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB28_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB28_4
-; GFX11-NEXT: .LBB28_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB28_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_2
-; GFX11-NEXT: .LBB28_4: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <2 x i16> %a, splat (i16 3)
- %a2 = bitcast <2 x i16> %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x i16> %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_v2i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_4
-; GCN-NEXT: .LBB29_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB29_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v4
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: .LBB29_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_v2i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB29_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB29_4
-; VI-NEXT: .LBB29_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB29_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: .LBB29_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_v2i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB29_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB29_4
-; GFX9-NEXT: .LBB29_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB29_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB29_2
-; GFX9-NEXT: .LBB29_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_v2i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB29_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB29_4
-; GFX11-NEXT: .LBB29_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB29_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_2
-; GFX11-NEXT: .LBB29_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to <2 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to <2 x i16>
- br label %end
-
-end:
- %phi = phi <2 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x i16> %phi
-}
-
-define <2 x bfloat> @bitcast_v2f16_to_v2bf16(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB30_4
-; GCN-NEXT: .LBB30_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB30_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v3
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_2
-; GCN-NEXT: .LBB30_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 0x200
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v1, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define <2 x half> @bitcast_v2bf16_to_v2f16(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB31_4
-; GCN-NEXT: .LBB31_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB31_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_2
-; GCN-NEXT: .LBB31_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB31_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v1, 16
-; VI-NEXT: .LBB31_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB31_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v1, s6
-; GFX9-NEXT: .LBB31_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB31_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
-; GFX11-NEXT: .LBB31_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define <1 x i32> @bitcast_v2f16_to_v1i32(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_4
-; GCN-NEXT: .LBB32_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB32_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_2
-; GCN-NEXT: .LBB32_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 0x200
-; VI-NEXT: v_add_f16_sdwa v1, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
-
-define <2 x half> @bitcast_v1i32_to_v2f16(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB33_4
-; GCN-NEXT: .LBB33_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB33_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_2
-; GCN-NEXT: .LBB33_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define <4 x i8> @bitcast_v2f16_to_v4i8(<2 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2f16_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB34_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB34_4
-; GCN-NEXT: .LBB34_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB34_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: v_bfe_u32 v3, v2, 8, 8
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_2
-; GCN-NEXT: .LBB34_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: v_bfe_u32 v3, v2, 8, 8
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2f16_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB34_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB34_4
-; VI-NEXT: .LBB34_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB34_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_2
-; VI-NEXT: .LBB34_4: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v1, 0x200
-; VI-NEXT: v_add_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v0, v1
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; VI-NEXT: v_bfe_u32 v3, v2, 8, 8
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2f16_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB34_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB34_4
-; GFX9-NEXT: .LBB34_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB34_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_2
-; GFX9-NEXT: .LBB34_4: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2f16_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB34_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB34_4
-; GFX11-NEXT: .LBB34_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB34_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_2
-; GFX11-NEXT: .LBB34_4: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <2 x half> %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x half> %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_v2f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB35_4
-; GCN-NEXT: .LBB35_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB35_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v4
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_2
-; GCN-NEXT: .LBB35_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v5
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v0, v3, v0
-; GCN-NEXT: v_or_b32_e32 v1, v4, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x300, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB35_4
-; VI-NEXT: .LBB35_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB35_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_2
-; VI-NEXT: .LBB35_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_v2f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB35_4
-; GFX9-NEXT: .LBB35_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB35_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_2
-; GFX9-NEXT: .LBB35_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_4
-; GFX11-NEXT: .LBB35_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB35_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB35_2
-; GFX11-NEXT: .LBB35_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to <2 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to <2 x half>
- br label %end
-
-end:
- %phi = phi <2 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x half> %phi
-}
-
-define <1 x i32> @bitcast_v2bf16_to_v1i32(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB36_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB36_4
-; GCN-NEXT: .LBB36_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB36_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_2
-; GCN-NEXT: .LBB36_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v1, 16
-; VI-NEXT: .LBB36_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v1, s6
-; GFX9-NEXT: .LBB36_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
-; GFX11-NEXT: .LBB36_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
-
-define <2 x bfloat> @bitcast_v1i32_to_v2bf16(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v2, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_4
-; GCN-NEXT: .LBB37_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB37_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_2
-; GCN-NEXT: .LBB37_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v1
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define <4 x i8> @bitcast_v2bf16_to_v4i8(<2 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v2bf16_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB38_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB38_4
-; GCN-NEXT: .LBB38_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB38_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v4
-; GCN-NEXT: v_alignbit_b32 v0, v2, v5, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_2
-; GCN-NEXT: .LBB38_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v2bf16_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB38_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB38_4
-; VI-NEXT: .LBB38_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB38_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB38_2
-; VI-NEXT: .LBB38_4: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_bfe_u32 v2, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v0
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_alignbit_b32 v4, v2, v1, 16
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v4
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v2bf16_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB38_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB38_4
-; GFX9-NEXT: .LBB38_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB38_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB38_2
-; GFX9-NEXT: .LBB38_4: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v2, v2, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v4, v0, v1, s6
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v4
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v2bf16_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB38_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB38_4
-; GFX11-NEXT: .LBB38_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB38_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB38_2
-; GFX11-NEXT: .LBB38_4: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
-; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v4, v0, v1, 0x7060302
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v4
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <2 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <2 x bfloat> %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <2 x bfloat> %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_v2bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB39_4
-; GCN-NEXT: .LBB39_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB39_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_2
-; GCN-NEXT: .LBB39_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v5
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v2, v3, v0
-; GCN-NEXT: v_or_b32_e32 v0, v4, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x3000000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x3000000, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB39_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB39_4
-; VI-NEXT: .LBB39_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB39_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB39_2
-; VI-NEXT: .LBB39_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_v2bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB39_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB39_4
-; GFX9-NEXT: .LBB39_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB39_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB39_2
-; GFX9-NEXT: .LBB39_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB39_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB39_4
-; GFX11-NEXT: .LBB39_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB39_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB39_2
-; GFX11-NEXT: .LBB39_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to <2 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to <2 x bfloat>
- br label %end
-
-end:
- %phi = phi <2 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <2 x bfloat> %phi
-}
-
-define <4 x i8> @bitcast_v1i32_to_v4i8(<1 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v1i32_to_v4i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB40_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB40_4
-; GCN-NEXT: .LBB40_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB40_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_2
-; GCN-NEXT: .LBB40_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v1i32_to_v4i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB40_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB40_4
-; VI-NEXT: .LBB40_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB40_3: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB40_2
-; VI-NEXT: .LBB40_4: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v1i32_to_v4i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr3
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB40_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB40_4
-; GFX9-NEXT: .LBB40_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB40_3: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB40_2
-; GFX9-NEXT: .LBB40_4: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX9-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v1i32_to_v4i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB40_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB40_4
-; GFX11-NEXT: .LBB40_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB40_3: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB40_2
-; GFX11-NEXT: .LBB40_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <1 x i32> %a, splat (i32 3)
- %a2 = bitcast <1 x i32> %a1 to <4 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <1 x i32> %a to <4 x i8>
- br label %end
-
-end:
- %phi = phi <4 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <4 x i8> %phi
-}
-
-define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v4i8_to_v1i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB41_4
-; GCN-NEXT: .LBB41_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB41_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_2
-; GCN-NEXT: .LBB41_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v0, v4, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v1, v1, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x3000000, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v4i8_to_v1i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v5, v0
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; VI-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB41_3
-; VI-NEXT: ; %bb.1: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB41_4
-; VI-NEXT: .LBB41_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB41_3: ; %cmp.false
-; VI-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr4
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB41_2
-; VI-NEXT: .LBB41_4: ; %cmp.true
-; VI-NEXT: v_add_u16_e32 v0, 3, v5
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_mov_b32_e32 v2, 0x300
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v4i8_to_v1i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v5, v0
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4
-; GFX9-NEXT: v_lshlrev_b16_e32 v4, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v3
-; GFX9-NEXT: ; implicit-def: $vgpr0
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB41_3
-; GFX9-NEXT: ; %bb.1: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execnz .LBB41_4
-; GFX9-NEXT: .LBB41_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB41_3: ; %cmp.false
-; GFX9-NEXT: v_or_b32_sdwa v0, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr5
-; GFX9-NEXT: ; implicit-def: $vgpr4
-; GFX9-NEXT: ; implicit-def: $vgpr2
-; GFX9-NEXT: ; implicit-def: $vgpr1
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB41_2
-; GFX9-NEXT: .LBB41_4: ; %cmp.true
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v5
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v4i8_to_v1i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v5, v0
-; GFX11-NEXT: v_lshlrev_b16 v1, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v3, 8, v3
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr0
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v4
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB41_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB41_4
-; GFX11-NEXT: .LBB41_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB41_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v5
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: ; implicit-def: $vgpr5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: ; implicit-def: $vgpr2
-; GFX11-NEXT: ; implicit-def: $vgpr3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB41_2
-; GFX11-NEXT: .LBB41_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v5, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <4 x i8> %a, splat (i8 3)
- %a2 = bitcast <4 x i8> %a1 to <1 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <4 x i8> %a to <1 x i32>
- br label %end
-
-end:
- %phi = phi <1 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <1 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.352bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.352bit.ll
deleted file mode 100644
index 922a47ea77fcd..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.352bit.ll
+++ /dev/null
@@ -1,228 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <11 x float> @bitcast_v11i32_to_v11f32(<11 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v11i32_to_v11f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v11i32_to_v11f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v11i32_to_v11f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v11i32_to_v11f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v11
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <11 x i32> %a, splat (i32 3)
- %a2 = bitcast <11 x i32> %a1 to <11 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <11 x i32> %a to <11 x float>
- br label %end
-
-end:
- %phi = phi <11 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <11 x float> %phi
-}
-
-define <11 x i32> @bitcast_v11f32_to_v11i32(<11 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v11f32_to_v11i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v11f32_to_v11i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v11f32_to_v11i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v11f32_to_v11i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v11
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v9, 1.0, v9
-; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v7, 1.0, v7
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <11 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <11 x float> %a1 to <11 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <11 x float> %a to <11 x i32>
- br label %end
-
-end:
- %phi = phi <11 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <11 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.384bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.384bit.ll
deleted file mode 100644
index f67af98fba0fa..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.384bit.ll
+++ /dev/null
@@ -1,1338 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <12 x float> @bitcast_v12i32_to_v12f32(<12 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12i32_to_v12f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12i32_to_v12f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12i32_to_v12f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12i32_to_v12f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <12 x i32> %a, splat (i32 3)
- %a2 = bitcast <12 x i32> %a1 to <12 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x i32> %a to <12 x float>
- br label %end
-
-end:
- %phi = phi <12 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x float> %phi
-}
-
-define <12 x i32> @bitcast_v12f32_to_v12i32(<12 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12f32_to_v12i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12f32_to_v12i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12f32_to_v12i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12f32_to_v12i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <12 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <12 x float> %a1 to <12 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x float> %a to <12 x i32>
- br label %end
-
-end:
- %phi = phi <12 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x i32> %phi
-}
-
-define <6 x double> @bitcast_v12i32_to_v6f64(<12 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12i32_to_v6f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12i32_to_v6f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12i32_to_v6f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12i32_to_v6f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <12 x i32> %a, splat (i32 3)
- %a2 = bitcast <12 x i32> %a1 to <6 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x i32> %a to <6 x double>
- br label %end
-
-end:
- %phi = phi <6 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x double> %phi
-}
-
-define <12 x i32> @bitcast_v6f64_to_v12i32(<6 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f64_to_v12i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f64_to_v12i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB3_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f64_to_v12i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB3_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB3_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f64_to_v12i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <6 x double> %a1 to <12 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x double> %a to <12 x i32>
- br label %end
-
-end:
- %phi = phi <12 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x i32> %phi
-}
-
-define <6 x i64> @bitcast_v12i32_to_v6i64(<12 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12i32_to_v6i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12i32_to_v6i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB4_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12i32_to_v6i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB4_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12i32_to_v6i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB4_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <12 x i32> %a, splat (i32 3)
- %a2 = bitcast <12 x i32> %a1 to <6 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x i32> %a to <6 x i64>
- br label %end
-
-end:
- %phi = phi <6 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i64> %phi
-}
-
-define <12 x i32> @bitcast_v6i64_to_v12i32(<6 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i64_to_v12i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i64_to_v12i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i64_to_v12i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i64_to_v12i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i64> %a, splat (i64 3)
- %a2 = bitcast <6 x i64> %a1 to <12 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i64> %a to <12 x i32>
- br label %end
-
-end:
- %phi = phi <12 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x i32> %phi
-}
-
-define <6 x double> @bitcast_v12f32_to_v6f64(<12 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12f32_to_v6f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB6_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12f32_to_v6f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB6_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB6_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12f32_to_v6f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB6_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB6_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12f32_to_v6f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <12 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <12 x float> %a1 to <6 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x float> %a to <6 x double>
- br label %end
-
-end:
- %phi = phi <6 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x double> %phi
-}
-
-define <12 x float> @bitcast_v6f64_to_v12f32(<6 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f64_to_v12f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f64_to_v12f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f64_to_v12f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB7_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f64_to_v12f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB7_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <6 x double> %a1 to <12 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x double> %a to <12 x float>
- br label %end
-
-end:
- %phi = phi <12 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x float> %phi
-}
-
-define <6 x i64> @bitcast_v12f32_to_v6i64(<12 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v12f32_to_v6i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB8_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v12f32_to_v6i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB8_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB8_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v12f32_to_v6i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB8_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB8_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v12f32_to_v6i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <12 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <12 x float> %a1 to <6 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <12 x float> %a to <6 x i64>
- br label %end
-
-end:
- %phi = phi <6 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i64> %phi
-}
-
-define <12 x float> @bitcast_v6i64_to_v12f32(<6 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i64_to_v12f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB9_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i64_to_v12f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i64_to_v12f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB9_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB9_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i64_to_v12f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i64> %a, splat (i64 3)
- %a2 = bitcast <6 x i64> %a1 to <12 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i64> %a to <12 x float>
- br label %end
-
-end:
- %phi = phi <12 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <12 x float> %phi
-}
-
-define <6 x i64> @bitcast_v6f64_to_v6i64(<6 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6f64_to_v6i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6f64_to_v6i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB10_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: .LBB10_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6f64_to_v6i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB10_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: .LBB10_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6f64_to_v6i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB10_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: .LBB10_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <6 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <6 x double> %a1 to <6 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x double> %a to <6 x i64>
- br label %end
-
-end:
- %phi = phi <6 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x i64> %phi
-}
-
-define <6 x double> @bitcast_v6i64_to_v6f64(<6 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v6i64_to_v6f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: .LBB11_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v6i64_to_v6f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v6i64_to_v6f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v6i64_to_v6f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v12
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <6 x i64> %a, splat (i64 3)
- %a2 = bitcast <6 x i64> %a1 to <6 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <6 x i64> %a to <6 x double>
- br label %end
-
-end:
- %phi = phi <6 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <6 x double> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.448bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.448bit.ll
deleted file mode 100644
index a5764f9da3194..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.448bit.ll
+++ /dev/null
@@ -1,714 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <14 x i32> @bitcast_v7i64_to_v14i32(<7 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7i64_to_v14i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7i64_to_v14i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7i64_to_v14i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7i64_to_v14i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <7 x i64> %a, splat (i64 3)
- %a2 = bitcast <7 x i64> %a1 to <14 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x i64> %a to <14 x i32>
- br label %end
-
-end:
- %phi = phi <14 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <14 x i32> %phi
-}
-
-define <7 x i64> @bitcast_v14i32_to_v7i64(<14 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v14i32_to_v7i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v14i32_to_v7i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v14i32_to_v7i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v14i32_to_v7i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB1_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB1_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <14 x i32> %a, splat (i32 3)
- %a2 = bitcast <14 x i32> %a1 to <7 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <14 x i32> %a to <7 x i64>
- br label %end
-
-end:
- %phi = phi <7 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x i64> %phi
-}
-
-define <7 x double> @bitcast_v7i64_to_v7f64(<7 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7i64_to_v7f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7i64_to_v7f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7i64_to_v7f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7i64_to_v7f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <7 x i64> %a, splat (i64 3)
- %a2 = bitcast <7 x i64> %a1 to <7 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x i64> %a to <7 x double>
- br label %end
-
-end:
- %phi = phi <7 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x double> %phi
-}
-
-define <7 x i64> @bitcast_v7f64_to_v7i64(<7 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7f64_to_v7i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7f64_to_v7i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: .LBB3_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7f64_to_v7i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB3_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: .LBB3_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7f64_to_v7i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <7 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <7 x double> %a1 to <7 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x double> %a to <7 x i64>
- br label %end
-
-end:
- %phi = phi <7 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x i64> %phi
-}
-
-define <7 x double> @bitcast_v14i32_to_v7f64(<14 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v14i32_to_v7f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v14i32_to_v7f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB4_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v14i32_to_v7f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB4_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v14i32_to_v7f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB4_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <14 x i32> %a, splat (i32 3)
- %a2 = bitcast <14 x i32> %a1 to <7 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <14 x i32> %a to <7 x double>
- br label %end
-
-end:
- %phi = phi <7 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <7 x double> %phi
-}
-
-define <14 x i32> @bitcast_v7f64_to_v14i32(<7 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v7f64_to_v14i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v7f64_to_v14i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v7f64_to_v14i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v7f64_to_v14i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v14
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <7 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <7 x double> %a1 to <14 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <7 x double> %a to <14 x i32>
- br label %end
-
-end:
- %phi = phi <14 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <14 x i32> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
deleted file mode 100644
index d09aaf12161a2..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
+++ /dev/null
@@ -1,775 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <3 x half> @bitcast_v3bf16_to_v3f16(<3 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3bf16_to_v3f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB0_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB0_4
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB0_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: .LBB0_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3bf16_to_v3f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_bfe_u32 v3, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v0
-; VI-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; VI-NEXT: v_mov_b32_e32 v2, 0x7fc00000
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3bf16_to_v3f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v3, v3, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v3, v3, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v2, s6
-; GFX9-NEXT: s_movk_i32 s6, 0x7fc0
-; GFX9-NEXT: v_alignbit_b32 v1, s6, v1, 16
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3bf16_to_v3f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v2, 16, v0
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v4, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_alignbit_b32 v1, 0x7fc0, v1, 16
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <3 x bfloat> %a1 to <3 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x bfloat> %a to <3 x half>
- br label %end
-
-end:
- %phi = phi <3 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x half> %phi
-}
-
-define <3 x bfloat> @bitcast_v3f16_to_v3bf16(<3 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3f16_to_v3bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB1_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB1_4
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB1_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: .LBB1_4: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3f16_to_v3bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v3, 0x200
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v2, v0
-; VI-NEXT: v_or_b32_e32 v1, 0x7e000000, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3f16_to_v3bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3f16_to_v3bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <3 x half> %a1 to <3 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x half> %a to <3 x bfloat>
- br label %end
-
-end:
- %phi = phi <3 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x bfloat> %phi
-}
-
-define <3 x i16> @bitcast_v3bf16_to_v3i16(<3 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3bf16_to_v3i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB2_4
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB2_3: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: .LBB2_4: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v5
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_alignbit_b32 v0, v3, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3bf16_to_v3i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_bfe_u32 v2, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; VI-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
-; VI-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v0
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_bfe_u32 v3, v2, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v2
-; VI-NEXT: v_add_u32_e32 v3, vcc, s6, v3
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; VI-NEXT: v_bfe_u32 v3, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v3, vcc, v3, v0
-; VI-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
-; VI-NEXT: v_or_b32_e32 v4, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v2, 16
-; VI-NEXT: v_mov_b32_e32 v2, 0x7fc00000
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3bf16_to_v3i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_add3_u32 v2, v2, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v3, v3, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
-; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v3, v3, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v0, v0, v2, s6
-; GFX9-NEXT: s_movk_i32 s6, 0x7fc0
-; GFX9-NEXT: v_alignbit_b32 v1, s6, v1, 16
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3bf16_to_v3i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v2, 16, v0
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v4, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v1
-; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_alignbit_b32 v1, 0x7fc0, v1, 16
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <3 x bfloat> %a1 to <3 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x bfloat> %a to <3 x i16>
- br label %end
-
-end:
- %phi = phi <3 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x i16> %phi
-}
-
-define <3 x bfloat> @bitcast_v3i16_to_v3bf16(<3 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3i16_to_v3bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v4, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB3_4
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB3_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: .LBB3_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v4
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3i16_to_v3bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 3
-; VI-NEXT: v_add_u16_sdwa v2, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_e32 v0, v0, v2
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3i16_to_v3bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3i16_to_v3bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <3 x i16> %a, splat (i16 3)
- %a2 = bitcast <3 x i16> %a1 to <3 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x i16> %a to <3 x bfloat>
- br label %end
-
-end:
- %phi = phi <3 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x bfloat> %phi
-}
-
-define <3 x i16> @bitcast_v3f16_to_v3i16(<3 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3f16_to_v3i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3f16_to_v3i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v3, 0x200
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v2, v0
-; VI-NEXT: v_or_b32_e32 v1, 0x7e000000, v1
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3f16_to_v3i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3f16_to_v3i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <3 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <3 x half> %a1 to <3 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x half> %a to <3 x i16>
- br label %end
-
-end:
- %phi = phi <3 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x i16> %phi
-}
-
-define <3 x half> @bitcast_v3i16_to_v3f16(<3 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v3i16_to_v3f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v6, v2
-; GCN-NEXT: v_mov_b32_e32 v4, v1
-; GCN-NEXT: v_mov_b32_e32 v5, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB5_4
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB5_3: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v6
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: .LBB5_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v3i16_to_v3f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v2, 3
-; VI-NEXT: v_add_u16_sdwa v2, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_or_b32_e32 v0, v0, v2
-; VI-NEXT: ; %bb.2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v3i16_to_v3f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v3i16_to_v3f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v2
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: ; %bb.2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <3 x i16> %a, splat (i16 3)
- %a2 = bitcast <3 x i16> %a1 to <3 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <3 x i16> %a to <3 x half>
- br label %end
-
-end:
- %phi = phi <3 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <3 x half> %phi
-}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
deleted file mode 100644
index f0ce1784eb107..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
+++ /dev/null
@@ -1,41662 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
-; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s
-
-define <16 x float> @bitcast_v16i32_to_v16f32(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB0_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB0_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB0_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB0_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB0_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB0_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB0_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB0_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <16 x i32> @bitcast_v16f32_to_v16i32(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB1_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB1_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB1_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB1_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB1_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB1_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB1_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB1_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <8 x i64> @bitcast_v16i32_to_v8i64(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB2_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB2_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB2_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB2_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB2_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB2_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <16 x i32> @bitcast_v8i64_to_v16i32(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB3_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB3_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB3_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB3_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB3_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB3_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <8 x double> @bitcast_v16i32_to_v8f64(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB4_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: .LBB4_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB4_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB4_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB4_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB4_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB4_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <16 x i32> @bitcast_v8f64_to_v16i32(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB5_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB5_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB5_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB5_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB5_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB5_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB5_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB5_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <32 x i16> @bitcast_v16i32_to_v32i16(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v30, v15
-; GCN-NEXT: v_mov_b32_e32 v28, v14
-; GCN-NEXT: v_mov_b32_e32 v26, v13
-; GCN-NEXT: v_mov_b32_e32 v24, v12
-; GCN-NEXT: v_mov_b32_e32 v22, v11
-; GCN-NEXT: v_mov_b32_e32 v20, v10
-; GCN-NEXT: v_mov_b32_e32 v18, v9
-; GCN-NEXT: v_mov_b32_e32 v32, v8
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v8, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB6_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB6_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v16, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB6_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB6_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB6_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB6_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB6_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB6_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <16 x i32> @bitcast_v32i16_to_v16i32(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB7_4
-; GCN-NEXT: .LBB7_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB7_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; GCN-NEXT: v_or_b32_e32 v0, v0, v54
-; GCN-NEXT: v_or_b32_e32 v1, v1, v55
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_or_b32_e32 v3, v3, v48
-; GCN-NEXT: v_or_b32_e32 v4, v4, v49
-; GCN-NEXT: v_or_b32_e32 v5, v5, v50
-; GCN-NEXT: v_or_b32_e32 v6, v6, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v52
-; GCN-NEXT: v_or_b32_e32 v8, v8, v17
-; GCN-NEXT: v_or_b32_e32 v9, v9, v19
-; GCN-NEXT: v_or_b32_e32 v10, v10, v21
-; GCN-NEXT: v_or_b32_e32 v11, v11, v23
-; GCN-NEXT: v_or_b32_e32 v12, v12, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v27
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_or_b32_e32 v15, v15, v53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: .LBB7_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v0, v54, v0
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_or_b32_e32 v4, v49, v4
-; GCN-NEXT: v_or_b32_e32 v5, v50, v5
-; GCN-NEXT: v_or_b32_e32 v6, v51, v6
-; GCN-NEXT: v_or_b32_e32 v7, v52, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v19, v9
-; GCN-NEXT: v_or_b32_e32 v10, v21, v10
-; GCN-NEXT: v_or_b32_e32 v11, v23, v11
-; GCN-NEXT: v_or_b32_e32 v12, v25, v12
-; GCN-NEXT: v_or_b32_e32 v13, v27, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_or_b32_e32 v15, v53, v15
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB7_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 3
-; VI-NEXT: v_add_u16_e32 v16, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: v_add_u16_e32 v16, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v16, v14
-; VI-NEXT: v_add_u16_e32 v16, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v16, v13
-; VI-NEXT: v_add_u16_e32 v16, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v16, v12
-; VI-NEXT: v_add_u16_e32 v16, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v16, v11
-; VI-NEXT: v_add_u16_e32 v16, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v16, v10
-; VI-NEXT: v_add_u16_e32 v16, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v16, v9
-; VI-NEXT: v_add_u16_e32 v16, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v16, v8
-; VI-NEXT: v_add_u16_e32 v16, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_add_u16_e32 v16, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v16, v6
-; VI-NEXT: v_add_u16_e32 v16, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v16, v5
-; VI-NEXT: v_add_u16_e32 v16, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v16, v4
-; VI-NEXT: v_add_u16_e32 v16, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v16, v3
-; VI-NEXT: v_add_u16_e32 v16, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v16, v2
-; VI-NEXT: v_add_u16_e32 v16, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v16, v1
-; VI-NEXT: v_add_u16_e32 v16, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB7_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB7_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB7_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB7_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <32 x half> @bitcast_v16i32_to_v32f16(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v33, v15
-; GCN-NEXT: v_mov_b32_e32 v34, v14
-; GCN-NEXT: v_mov_b32_e32 v35, v13
-; GCN-NEXT: v_mov_b32_e32 v36, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v11
-; GCN-NEXT: v_mov_b32_e32 v38, v10
-; GCN-NEXT: v_mov_b32_e32 v39, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v49, v7
-; GCN-NEXT: v_mov_b32_e32 v50, v6
-; GCN-NEXT: v_mov_b32_e32 v51, v5
-; GCN-NEXT: v_mov_b32_e32 v52, v4
-; GCN-NEXT: v_mov_b32_e32 v53, v3
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v49
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v50
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v51
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v52
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v53
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v54
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: .LBB8_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB8_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v55
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32
-; GCN-NEXT: .LBB8_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB8_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB8_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB8_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB8_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v32f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB8_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB8_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <32 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <32 x half>
- br label %end
-
-end:
- %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x half> %phi
-}
-
-define <16 x i32> @bitcast_v32f16_to_v16i32(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43
-; GCN-NEXT: v_or_b32_e32 v0, v44, v0
-; GCN-NEXT: v_or_b32_e32 v1, v42, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_or_b32_e32 v3, v50, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v36, v6
-; GCN-NEXT: v_or_b32_e32 v7, v34, v7
-; GCN-NEXT: v_or_b32_e32 v8, v33, v8
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: v_or_b32_e32 v10, v31, v10
-; GCN-NEXT: v_or_b32_e32 v11, v21, v11
-; GCN-NEXT: v_or_b32_e32 v12, v19, v12
-; GCN-NEXT: v_or_b32_e32 v13, v18, v13
-; GCN-NEXT: v_or_b32_e32 v14, v17, v14
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB9_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB9_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v26, v24
-; GCN-NEXT: v_or_b32_e32 v10, v28, v27
-; GCN-NEXT: v_or_b32_e32 v11, v21, v29
-; GCN-NEXT: v_or_b32_e32 v12, v19, v25
-; GCN-NEXT: v_or_b32_e32 v13, v18, v23
-; GCN-NEXT: v_or_b32_e32 v14, v17, v22
-; GCN-NEXT: v_or_b32_e32 v15, v16, v20
-; GCN-NEXT: .LBB9_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB9_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 0x200
-; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v17
-; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v17
-; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v17
-; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v17
-; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v17
-; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v17
-; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v17
-; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v17
-; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v17
-; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v17
-; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v17
-; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v17
-; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v17
-; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v17
-; VI-NEXT: v_or_b32_e32 v0, v0, v16
-; VI-NEXT: .LBB9_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB9_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB9_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB9_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <32 x bfloat> @bitcast_v16i32_to_v32bf16(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v55, v15
-; GCN-NEXT: v_mov_b32_e32 v54, v14
-; GCN-NEXT: v_mov_b32_e32 v53, v13
-; GCN-NEXT: v_mov_b32_e32 v52, v12
-; GCN-NEXT: v_mov_b32_e32 v51, v11
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v49, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v39, v7
-; GCN-NEXT: v_mov_b32_e32 v38, v6
-; GCN-NEXT: v_mov_b32_e32 v37, v5
-; GCN-NEXT: v_mov_b32_e32 v36, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v3
-; GCN-NEXT: v_mov_b32_e32 v34, v2
-; GCN-NEXT: v_mov_b32_e32 v33, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB10_4
-; GCN-NEXT: .LBB10_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB10_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB10_2
-; GCN-NEXT: .LBB10_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v55
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB10_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: .LBB10_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB10_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: .LBB10_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB10_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
-; GFX11-NEXT: .LBB10_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <16 x i32> @bitcast_v32bf16_to_v16i32(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42
-; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16
-; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16
-; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16
-; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB11_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB11_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16
-; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16
-; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16
-; GCN-NEXT: .LBB11_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB11_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB11_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB11_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7
-; GFX9-NEXT: .LBB11_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32bf16_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB11_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19
-; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302
-; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23
-; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302
-; GFX11-NEXT: .LBB11_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <32 x bfloat> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x bfloat> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16i32_to_v64i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB12_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB12_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB12_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_or_b32_e32 v60, v1, v18
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17
-; GCN-NEXT: v_or_b32_e32 v17, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46
-; GCN-NEXT: v_or_b32_e32 v46, v1, v2
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v7, v2, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47
-; GCN-NEXT: v_or_b32_e32 v54, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48
-; GCN-NEXT: v_or_b32_e32 v9, v4, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v42
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55
-; GCN-NEXT: v_or_b32_e32 v48, v5, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34
-; GCN-NEXT: v_or_b32_e32 v34, v6, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v50
-; GCN-NEXT: v_or_b32_e32 v11, v8, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28
-; GCN-NEXT: v_or_b32_e32 v13, v10, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v38
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35
-; GCN-NEXT: v_or_b32_e32 v28, v12, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 24, v24
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21
-; GCN-NEXT: v_or_b32_e32 v21, v14, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v31
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29
-; GCN-NEXT: v_or_b32_e32 v29, v15, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v31, v19, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45
-; GCN-NEXT: v_or_b32_e32 v27, v61, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41
-; GCN-NEXT: v_or_b32_e32 v40, v40, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57
-; GCN-NEXT: v_or_b32_e32 v56, v56, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46
-; GCN-NEXT: v_or_b32_e32 v49, v49, v47
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44
-; GCN-NEXT: v_or_b32_e32 v51, v51, v43
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v36, v36, v55
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v37, v37, v53
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v30, v30, v50
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v32, v32, v39
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GCN-NEXT: v_or_b32_e32 v24, v24, v35
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v25, v25, v38
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v26, v26, v42
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v33, v33, v45
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v23, v41, v23
-; GCN-NEXT: v_or_b32_e32 v31, v57, v31
-; GCN-NEXT: v_or_b32_e32 v17, v17, v27
-; GCN-NEXT: v_or_b32_e32 v27, v58, v40
-; GCN-NEXT: v_or_b32_e32 v35, v59, v56
-; GCN-NEXT: v_or_b32_e32 v38, v46, v49
-; GCN-NEXT: v_or_b32_e32 v39, v44, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v36
-; GCN-NEXT: v_or_b32_e32 v36, v54, v37
-; GCN-NEXT: v_or_b32_e32 v9, v9, v30
-; GCN-NEXT: v_or_b32_e32 v30, v48, v32
-; GCN-NEXT: v_or_b32_e32 v24, v34, v24
-; GCN-NEXT: v_or_b32_e32 v11, v11, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v20
-; GCN-NEXT: v_or_b32_e32 v20, v28, v26
-; GCN-NEXT: v_or_b32_e32 v21, v21, v33
-; GCN-NEXT: v_or_b32_e32 v23, v29, v23
-; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16i32_to_v64i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_mov_b32_e32 v26, v22
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: .LBB12_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB12_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: .LBB12_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16i32_to_v64i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_mov_b32_e32 v26, v23
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: .LBB12_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB12_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
-; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
-; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
-; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
-; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
-; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
-; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
-; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
-; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB12_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16i32_to_v64i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB12_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB12_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
-; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
-; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
-; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
-; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
-; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
-; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
-; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
-; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
-; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
-; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
-; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
-; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
-; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB12_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v24, v96, v24
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v86, v85
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v24
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v84
-; GFX11-NEXT: v_or_b32_e32 v23, v83, v23
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v80
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v71
-; GFX11-NEXT: v_or_b32_e32 v22, v70, v22
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v24
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v23
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v25
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-NEXT: v_or_b32_e32 v21, v23, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v22, v25, v54
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v24
-; GFX11-NEXT: v_or_b32_e32 v23, v25, v49
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v48
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v22
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69
-; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v19, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v22
-; GFX11-NEXT: v_or_b32_e32 v18, v23, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v69
-; GFX11-NEXT: v_or_b32_e32 v67, v68, v67
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v22
-; GFX11-NEXT: v_or_b32_e32 v17, v23, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v20
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v21
-; GFX11-NEXT: s_clause 0x3
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <16 x i32> %a, splat (i32 3)
- %a2 = bitcast <16 x i32> %a1 to <64 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x i32> %a to <64 x i8>
- br label %end
-
-end:
- %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i8> %phi
-}
-
-define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i8_to_v16i32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v34, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v33, v8
-; GCN-NEXT: v_mov_b32_e32 v36, v6
-; GCN-NEXT: v_mov_b32_e32 v32, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_or_b32_e32 v0, v0, v42
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_or_b32_e32 v1, v1, v41
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v40
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_or_b32_e32 v3, v3, v55
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32
-; GCN-NEXT: v_or_b32_e32 v17, v33, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v28, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
-; GCN-NEXT: v_or_b32_e32 v27, v35, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v29, v36, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37
-; GCN-NEXT: v_or_b32_e32 v31, v38, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v9, v9, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v15, v15, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v25, v59, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v30, v57, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v32, v58, v8
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v8, v10
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v8, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v8, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v8, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v19, v46, v19
-; GCN-NEXT: v_or_b32_e32 v0, v0, v7
-; GCN-NEXT: v_or_b32_e32 v1, v1, v6
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_or_b32_e32 v4, v17, v21
-; GCN-NEXT: v_or_b32_e32 v5, v20, v22
-; GCN-NEXT: v_or_b32_e32 v6, v23, v24
-; GCN-NEXT: v_or_b32_e32 v7, v26, v28
-; GCN-NEXT: v_or_b32_e32 v8, v27, v25
-; GCN-NEXT: v_or_b32_e32 v9, v29, v30
-; GCN-NEXT: v_or_b32_e32 v10, v31, v32
-; GCN-NEXT: v_or_b32_e32 v11, v33, v34
-; GCN-NEXT: v_or_b32_e32 v12, v35, v12
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v16
-; GCN-NEXT: v_or_b32_e32 v15, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; kill: killed $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: .LBB13_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB13_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v42, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v41, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v40, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v55, v3
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7
-; GCN-NEXT: v_or_b32_e32 v6, v17, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v27, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v17, v44, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22
-; GCN-NEXT: v_or_b32_e32 v20, v45, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v56, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32
-; GCN-NEXT: v_or_b32_e32 v29, v43, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v19, v47, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_or_b32_e32 v16, v59, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_or_b32_e32 v18, v57, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_or_b32_e32 v22, v58, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19
-; GCN-NEXT: v_or_b32_e32 v30, v46, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v0, v31, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v10, v9
-; GCN-NEXT: v_or_b32_e32 v6, v12, v11
-; GCN-NEXT: v_or_b32_e32 v7, v14, v13
-; GCN-NEXT: v_or_b32_e32 v8, v16, v15
-; GCN-NEXT: v_or_b32_e32 v9, v18, v17
-; GCN-NEXT: v_or_b32_e32 v10, v22, v20
-; GCN-NEXT: v_or_b32_e32 v11, v24, v23
-; GCN-NEXT: v_or_b32_e32 v12, v26, v25
-; GCN-NEXT: v_or_b32_e32 v13, v28, v27
-; GCN-NEXT: v_or_b32_e32 v14, v21, v29
-; GCN-NEXT: v_or_b32_e32 v15, v30, v19
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15
-; GCN-NEXT: .LBB13_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i8_to_v16i32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: .LBB13_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB13_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v15, 0x300
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v51
-; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v10, 3, v24
-; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v30
-; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v60
-; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v42
-; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v53
-; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v52
-; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v50
-; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v63
-; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v48
-; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v44
-; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v39
-; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v16
-; VI-NEXT: v_add_u16_e32 v16, 3, v19
-; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: .LBB13_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i8_to_v16i32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: .LBB13_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB13_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v52
-; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v50
-; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: .LBB13_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i8_to_v16i32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v14, off, s32
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(25)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: s_waitcnt vmcnt(21)
-; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65
-; GFX11-NEXT: s_waitcnt vmcnt(19)
-; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67
-; GFX11-NEXT: s_waitcnt vmcnt(17)
-; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: s_waitcnt vmcnt(15)
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(13)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(11)
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_4
-; GFX11-NEXT: .LBB13_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB13_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v118
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v119
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v117
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v6, v113
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v114
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v116
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v101
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v102
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v103
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v112
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v100
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v11
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v87
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v96
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v97
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v98
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v99
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v81
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v82
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v83
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v84
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v85
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v27
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v29
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v65
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v66
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v67
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v17, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v18, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v19, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v20, v24, v25
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB13_2
-; GFX11-NEXT: .LBB13_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v118, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v119, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v2, v117, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v114, v3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v4, v115, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v116, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v113, v6
-; GFX11-NEXT: v_add_nc_u16 v7, v16, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v6, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v101, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v100, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v102, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v103, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v112, v10
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v6, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v10, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v87, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v96, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v97, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v98, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v99, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v81, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v82, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v83, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v84, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v85, v15
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_add_nc_u16 v11, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v11, v27, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v29, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v65, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v66, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v67, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v25, v24
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i8> %a, splat (i8 3)
- %a2 = bitcast <64 x i8> %a1 to <16 x i32>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i8> %a to <16 x i32>
- br label %end
-
-end:
- %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x i32> %phi
-}
-
-define <8 x i64> @bitcast_v16f32_to_v8i64(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB14_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB14_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB14_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB14_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB14_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB14_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB14_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB14_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <16 x float> @bitcast_v8i64_to_v16f32(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB15_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: .LBB15_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB15_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB15_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB15_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB15_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB15_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB15_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <8 x double> @bitcast_v16f32_to_v8f64(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB16_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: .LBB16_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB16_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB16_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB16_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB16_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB16_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB16_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <16 x float> @bitcast_v8f64_to_v16f32(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB17_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: .LBB17_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB17_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB17_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB17_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB17_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB17_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB17_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <32 x i16> @bitcast_v16f32_to_v32i16(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v30, v15
-; GCN-NEXT: v_mov_b32_e32 v28, v14
-; GCN-NEXT: v_mov_b32_e32 v26, v13
-; GCN-NEXT: v_mov_b32_e32 v24, v12
-; GCN-NEXT: v_mov_b32_e32 v22, v11
-; GCN-NEXT: v_mov_b32_e32 v20, v10
-; GCN-NEXT: v_mov_b32_e32 v18, v9
-; GCN-NEXT: v_mov_b32_e32 v32, v8
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v8, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB18_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18
-; GCN-NEXT: v_add_f32_e32 v32, 1.0, v32
-; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22
-; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20
-; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26
-; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24
-; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30
-; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB18_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v16, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB18_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB18_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB18_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB18_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB18_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB18_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <16 x float> @bitcast_v32i16_to_v16f32(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB19_4
-; GCN-NEXT: .LBB19_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB19_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; GCN-NEXT: v_or_b32_e32 v0, v0, v54
-; GCN-NEXT: v_or_b32_e32 v1, v1, v55
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_or_b32_e32 v3, v3, v48
-; GCN-NEXT: v_or_b32_e32 v4, v4, v49
-; GCN-NEXT: v_or_b32_e32 v5, v5, v50
-; GCN-NEXT: v_or_b32_e32 v6, v6, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v52
-; GCN-NEXT: v_or_b32_e32 v8, v8, v17
-; GCN-NEXT: v_or_b32_e32 v9, v9, v19
-; GCN-NEXT: v_or_b32_e32 v10, v10, v21
-; GCN-NEXT: v_or_b32_e32 v11, v11, v23
-; GCN-NEXT: v_or_b32_e32 v12, v12, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v27
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_or_b32_e32 v15, v15, v53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB19_2
-; GCN-NEXT: .LBB19_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v0, v54, v0
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_or_b32_e32 v4, v49, v4
-; GCN-NEXT: v_or_b32_e32 v5, v50, v5
-; GCN-NEXT: v_or_b32_e32 v6, v51, v6
-; GCN-NEXT: v_or_b32_e32 v7, v52, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v19, v9
-; GCN-NEXT: v_or_b32_e32 v10, v21, v10
-; GCN-NEXT: v_or_b32_e32 v11, v23, v11
-; GCN-NEXT: v_or_b32_e32 v12, v25, v12
-; GCN-NEXT: v_or_b32_e32 v13, v27, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_or_b32_e32 v15, v53, v15
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB19_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 3
-; VI-NEXT: v_add_u16_e32 v16, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: v_add_u16_e32 v16, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v16, v14
-; VI-NEXT: v_add_u16_e32 v16, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v16, v13
-; VI-NEXT: v_add_u16_e32 v16, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v16, v12
-; VI-NEXT: v_add_u16_e32 v16, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v16, v11
-; VI-NEXT: v_add_u16_e32 v16, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v16, v10
-; VI-NEXT: v_add_u16_e32 v16, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v16, v9
-; VI-NEXT: v_add_u16_e32 v16, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v16, v8
-; VI-NEXT: v_add_u16_e32 v16, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_add_u16_e32 v16, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v16, v6
-; VI-NEXT: v_add_u16_e32 v16, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v16, v5
-; VI-NEXT: v_add_u16_e32 v16, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v16, v4
-; VI-NEXT: v_add_u16_e32 v16, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v16, v3
-; VI-NEXT: v_add_u16_e32 v16, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v16, v2
-; VI-NEXT: v_add_u16_e32 v16, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v16, v1
-; VI-NEXT: v_add_u16_e32 v16, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB19_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB19_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB19_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB19_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB19_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <32 x half> @bitcast_v16f32_to_v32f16(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v33, v15
-; GCN-NEXT: v_mov_b32_e32 v34, v14
-; GCN-NEXT: v_mov_b32_e32 v35, v13
-; GCN-NEXT: v_mov_b32_e32 v36, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v11
-; GCN-NEXT: v_mov_b32_e32 v38, v10
-; GCN-NEXT: v_mov_b32_e32 v39, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v49, v7
-; GCN-NEXT: v_mov_b32_e32 v50, v6
-; GCN-NEXT: v_mov_b32_e32 v51, v5
-; GCN-NEXT: v_mov_b32_e32 v52, v4
-; GCN-NEXT: v_mov_b32_e32 v53, v3
-; GCN-NEXT: v_mov_b32_e32 v54, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v49
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v50
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v51
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v52
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v53
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v54
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: .LBB20_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB20_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v32
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v55
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v54
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v53
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v52
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v51
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v50
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v49
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v48
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v39
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v38
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v37
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v36
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v35
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v34
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32
-; GCN-NEXT: .LBB20_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB20_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB20_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB20_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB20_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v32f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB20_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB20_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <32 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <32 x half>
- br label %end
-
-end:
- %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x half> %phi
-}
-
-define <16 x float> @bitcast_v32f16_to_v16f32(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43
-; GCN-NEXT: v_or_b32_e32 v0, v44, v0
-; GCN-NEXT: v_or_b32_e32 v1, v42, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_or_b32_e32 v3, v50, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v36, v6
-; GCN-NEXT: v_or_b32_e32 v7, v34, v7
-; GCN-NEXT: v_or_b32_e32 v8, v33, v8
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: v_or_b32_e32 v10, v31, v10
-; GCN-NEXT: v_or_b32_e32 v11, v21, v11
-; GCN-NEXT: v_or_b32_e32 v12, v19, v12
-; GCN-NEXT: v_or_b32_e32 v13, v18, v13
-; GCN-NEXT: v_or_b32_e32 v14, v17, v14
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB21_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB21_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v26, v24
-; GCN-NEXT: v_or_b32_e32 v10, v28, v27
-; GCN-NEXT: v_or_b32_e32 v11, v21, v29
-; GCN-NEXT: v_or_b32_e32 v12, v19, v25
-; GCN-NEXT: v_or_b32_e32 v13, v18, v23
-; GCN-NEXT: v_or_b32_e32 v14, v17, v22
-; GCN-NEXT: v_or_b32_e32 v15, v16, v20
-; GCN-NEXT: .LBB21_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB21_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 0x200
-; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v17
-; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v17
-; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v17
-; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v17
-; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v17
-; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v17
-; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v17
-; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v17
-; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v17
-; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v17
-; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v17
-; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v17
-; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v17
-; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v17
-; VI-NEXT: v_or_b32_e32 v0, v0, v16
-; VI-NEXT: .LBB21_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB21_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB21_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB21_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB21_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <32 x bfloat> @bitcast_v16f32_to_v32bf16(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v55, v15
-; GCN-NEXT: v_mov_b32_e32 v54, v14
-; GCN-NEXT: v_mov_b32_e32 v53, v13
-; GCN-NEXT: v_mov_b32_e32 v52, v12
-; GCN-NEXT: v_mov_b32_e32 v51, v11
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v49, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v39, v7
-; GCN-NEXT: v_mov_b32_e32 v38, v6
-; GCN-NEXT: v_mov_b32_e32 v37, v5
-; GCN-NEXT: v_mov_b32_e32 v36, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v3
-; GCN-NEXT: v_mov_b32_e32 v34, v2
-; GCN-NEXT: v_mov_b32_e32 v33, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB22_4
-; GCN-NEXT: .LBB22_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB22_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB22_2
-; GCN-NEXT: .LBB22_4: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v0, 1.0, v32
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v33
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v34
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v35
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v36
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v37
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v38
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v39
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v48
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v49
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v50
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v51
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v52
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v53
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v54
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v55
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB22_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
-; VI-NEXT: .LBB22_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB22_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
-; GFX9-NEXT: .LBB22_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB22_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-NEXT: .LBB22_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <16 x float> @bitcast_v32bf16_to_v16f32(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42
-; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16
-; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16
-; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16
-; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB23_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB23_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16
-; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16
-; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16
-; GCN-NEXT: .LBB23_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB23_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB23_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB23_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7
-; GFX9-NEXT: .LBB23_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32bf16_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB23_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19
-; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302
-; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23
-; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302
-; GFX11-NEXT: .LBB23_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <32 x bfloat> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x bfloat> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) {
-; GCN-LABEL: bitcast_v16f32_to_v64i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB24_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB24_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB24_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_or_b32_e32 v60, v1, v18
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17
-; GCN-NEXT: v_or_b32_e32 v17, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46
-; GCN-NEXT: v_or_b32_e32 v46, v1, v2
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v7, v2, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47
-; GCN-NEXT: v_or_b32_e32 v54, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48
-; GCN-NEXT: v_or_b32_e32 v9, v4, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v42
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55
-; GCN-NEXT: v_or_b32_e32 v48, v5, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34
-; GCN-NEXT: v_or_b32_e32 v34, v6, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v50
-; GCN-NEXT: v_or_b32_e32 v11, v8, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28
-; GCN-NEXT: v_or_b32_e32 v13, v10, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v38
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35
-; GCN-NEXT: v_or_b32_e32 v28, v12, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 24, v24
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21
-; GCN-NEXT: v_or_b32_e32 v21, v14, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v31
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29
-; GCN-NEXT: v_or_b32_e32 v29, v15, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v31, v19, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45
-; GCN-NEXT: v_or_b32_e32 v27, v61, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41
-; GCN-NEXT: v_or_b32_e32 v40, v40, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57
-; GCN-NEXT: v_or_b32_e32 v56, v56, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46
-; GCN-NEXT: v_or_b32_e32 v49, v49, v47
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44
-; GCN-NEXT: v_or_b32_e32 v51, v51, v43
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v36, v36, v55
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v37, v37, v53
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v30, v30, v50
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v32, v32, v39
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GCN-NEXT: v_or_b32_e32 v24, v24, v35
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v25, v25, v38
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v26, v26, v42
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v33, v33, v45
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v23, v41, v23
-; GCN-NEXT: v_or_b32_e32 v31, v57, v31
-; GCN-NEXT: v_or_b32_e32 v17, v17, v27
-; GCN-NEXT: v_or_b32_e32 v27, v58, v40
-; GCN-NEXT: v_or_b32_e32 v35, v59, v56
-; GCN-NEXT: v_or_b32_e32 v38, v46, v49
-; GCN-NEXT: v_or_b32_e32 v39, v44, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v36
-; GCN-NEXT: v_or_b32_e32 v36, v54, v37
-; GCN-NEXT: v_or_b32_e32 v9, v9, v30
-; GCN-NEXT: v_or_b32_e32 v30, v48, v32
-; GCN-NEXT: v_or_b32_e32 v24, v34, v24
-; GCN-NEXT: v_or_b32_e32 v11, v11, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v20
-; GCN-NEXT: v_or_b32_e32 v20, v28, v26
-; GCN-NEXT: v_or_b32_e32 v21, v21, v33
-; GCN-NEXT: v_or_b32_e32 v23, v29, v23
-; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v16f32_to_v64i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_mov_b32_e32 v26, v22
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: .LBB24_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB24_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
-; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
-; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
-; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
-; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
-; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
-; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
-; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
-; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
-; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
-; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: .LBB24_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v16f32_to_v64i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_mov_b32_e32 v26, v23
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: .LBB24_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB24_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
-; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
-; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
-; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
-; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
-; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
-; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
-; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
-; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB24_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v16f32_to_v64i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB24_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB24_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-NEXT: v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v7, 1.0, v7
-; GFX11-NEXT: v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v13, 1.0, v13
-; GFX11-NEXT: v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v9, 1.0, v9
-; GFX11-NEXT: v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v11, 1.0, v11
-; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB24_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v24, v96, v24
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v86, v85
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v24
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v84
-; GFX11-NEXT: v_or_b32_e32 v23, v83, v23
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v80
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v71
-; GFX11-NEXT: v_or_b32_e32 v22, v70, v22
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v24
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v23
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v25
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-NEXT: v_or_b32_e32 v21, v23, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v22, v25, v54
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v24
-; GFX11-NEXT: v_or_b32_e32 v23, v25, v49
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v48
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v22
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69
-; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v19, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v22
-; GFX11-NEXT: v_or_b32_e32 v18, v23, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v69
-; GFX11-NEXT: v_or_b32_e32 v67, v68, v67
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v22
-; GFX11-NEXT: v_or_b32_e32 v17, v23, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v20
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v21
-; GFX11-NEXT: s_clause 0x3
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <16 x float> %a, splat (float 1.000000e+00)
- %a2 = bitcast <16 x float> %a1 to <64 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <16 x float> %a to <64 x i8>
- br label %end
-
-end:
- %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i8> %phi
-}
-
-define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i8_to_v16f32:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v34, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v33, v8
-; GCN-NEXT: v_mov_b32_e32 v36, v6
-; GCN-NEXT: v_mov_b32_e32 v32, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_or_b32_e32 v0, v0, v42
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_or_b32_e32 v1, v1, v41
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v40
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_or_b32_e32 v3, v3, v55
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32
-; GCN-NEXT: v_or_b32_e32 v17, v33, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v28, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
-; GCN-NEXT: v_or_b32_e32 v27, v35, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v29, v36, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37
-; GCN-NEXT: v_or_b32_e32 v31, v38, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v9, v9, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v15, v15, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v25, v59, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v30, v57, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v32, v58, v8
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v8, v10
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v8, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v8, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v8, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v19, v46, v19
-; GCN-NEXT: v_or_b32_e32 v0, v0, v7
-; GCN-NEXT: v_or_b32_e32 v1, v1, v6
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_or_b32_e32 v4, v17, v21
-; GCN-NEXT: v_or_b32_e32 v5, v20, v22
-; GCN-NEXT: v_or_b32_e32 v6, v23, v24
-; GCN-NEXT: v_or_b32_e32 v7, v26, v28
-; GCN-NEXT: v_or_b32_e32 v8, v27, v25
-; GCN-NEXT: v_or_b32_e32 v9, v29, v30
-; GCN-NEXT: v_or_b32_e32 v10, v31, v32
-; GCN-NEXT: v_or_b32_e32 v11, v33, v34
-; GCN-NEXT: v_or_b32_e32 v12, v35, v12
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v16
-; GCN-NEXT: v_or_b32_e32 v15, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; kill: killed $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: .LBB25_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB25_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v42, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v41, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v40, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v55, v3
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7
-; GCN-NEXT: v_or_b32_e32 v6, v17, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v27, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v17, v44, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22
-; GCN-NEXT: v_or_b32_e32 v20, v45, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v56, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32
-; GCN-NEXT: v_or_b32_e32 v29, v43, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v19, v47, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_or_b32_e32 v16, v59, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_or_b32_e32 v18, v57, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_or_b32_e32 v22, v58, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19
-; GCN-NEXT: v_or_b32_e32 v30, v46, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v0, v31, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v10, v9
-; GCN-NEXT: v_or_b32_e32 v6, v12, v11
-; GCN-NEXT: v_or_b32_e32 v7, v14, v13
-; GCN-NEXT: v_or_b32_e32 v8, v16, v15
-; GCN-NEXT: v_or_b32_e32 v9, v18, v17
-; GCN-NEXT: v_or_b32_e32 v10, v22, v20
-; GCN-NEXT: v_or_b32_e32 v11, v24, v23
-; GCN-NEXT: v_or_b32_e32 v12, v26, v25
-; GCN-NEXT: v_or_b32_e32 v13, v28, v27
-; GCN-NEXT: v_or_b32_e32 v14, v21, v29
-; GCN-NEXT: v_or_b32_e32 v15, v30, v19
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15
-; GCN-NEXT: .LBB25_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i8_to_v16f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: .LBB25_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB25_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v15, 0x300
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v51
-; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v10, 3, v24
-; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v30
-; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v60
-; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v42
-; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v53
-; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v52
-; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v50
-; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v63
-; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v48
-; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v44
-; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v39
-; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v16
-; VI-NEXT: v_add_u16_e32 v16, 3, v19
-; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: .LBB25_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i8_to_v16f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: .LBB25_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB25_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v52
-; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v50
-; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: .LBB25_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i8_to_v16f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v14, off, s32
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(25)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: s_waitcnt vmcnt(21)
-; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65
-; GFX11-NEXT: s_waitcnt vmcnt(19)
-; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67
-; GFX11-NEXT: s_waitcnt vmcnt(17)
-; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: s_waitcnt vmcnt(15)
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(13)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(11)
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB25_4
-; GFX11-NEXT: .LBB25_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB25_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v118
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v119
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v117
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v6, v113
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v114
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v116
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v101
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v102
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v103
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v112
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v100
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v11
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v87
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v96
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v97
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v98
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v99
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v81
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v82
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v83
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v84
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v85
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v27
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v29
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v65
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v66
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v67
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v17, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v18, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v19, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v20, v24, v25
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB25_2
-; GFX11-NEXT: .LBB25_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v118, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v119, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v2, v117, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v114, v3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v4, v115, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v116, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v113, v6
-; GFX11-NEXT: v_add_nc_u16 v7, v16, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v6, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v101, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v100, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v102, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v103, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v112, v10
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v6, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v10, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v87, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v96, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v97, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v98, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v99, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v81, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v82, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v83, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v84, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v85, v15
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_add_nc_u16 v11, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v11, v27, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v29, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v65, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v66, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v67, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v25, v24
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i8> %a, splat (i8 3)
- %a2 = bitcast <64 x i8> %a1 to <16 x float>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i8> %a to <16 x float>
- br label %end
-
-end:
- %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <16 x float> %phi
-}
-
-define <8 x double> @bitcast_v8i64_to_v8f64(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB26_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; GCN-NEXT: .LBB26_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB26_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: .LBB26_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB26_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: .LBB26_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB26_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: .LBB26_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <8 x i64> @bitcast_v8f64_to_v8i64(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB27_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: .LBB27_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB27_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: .LBB27_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB27_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: .LBB27_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB27_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: .LBB27_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <32 x i16> @bitcast_v8i64_to_v32i16(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v30, v15
-; GCN-NEXT: v_mov_b32_e32 v28, v14
-; GCN-NEXT: v_mov_b32_e32 v26, v13
-; GCN-NEXT: v_mov_b32_e32 v24, v12
-; GCN-NEXT: v_mov_b32_e32 v22, v11
-; GCN-NEXT: v_mov_b32_e32 v20, v10
-; GCN-NEXT: v_mov_b32_e32 v18, v9
-; GCN-NEXT: v_mov_b32_e32 v32, v8
-; GCN-NEXT: v_mov_b32_e32 v14, v7
-; GCN-NEXT: v_mov_b32_e32 v12, v6
-; GCN-NEXT: v_mov_b32_e32 v10, v5
-; GCN-NEXT: v_mov_b32_e32 v8, v4
-; GCN-NEXT: v_mov_b32_e32 v6, v3
-; GCN-NEXT: v_mov_b32_e32 v4, v2
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB28_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32
-; GCN-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24
-; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28
-; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc
-; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; GCN-NEXT: .LBB28_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v16, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB28_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB28_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB28_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB28_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB28_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB28_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <8 x i64> @bitcast_v32i16_to_v8i64(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB29_4
-; GCN-NEXT: .LBB29_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB29_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; GCN-NEXT: v_or_b32_e32 v0, v0, v54
-; GCN-NEXT: v_or_b32_e32 v1, v1, v55
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_or_b32_e32 v3, v3, v48
-; GCN-NEXT: v_or_b32_e32 v4, v4, v49
-; GCN-NEXT: v_or_b32_e32 v5, v5, v50
-; GCN-NEXT: v_or_b32_e32 v6, v6, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v52
-; GCN-NEXT: v_or_b32_e32 v8, v8, v17
-; GCN-NEXT: v_or_b32_e32 v9, v9, v19
-; GCN-NEXT: v_or_b32_e32 v10, v10, v21
-; GCN-NEXT: v_or_b32_e32 v11, v11, v23
-; GCN-NEXT: v_or_b32_e32 v12, v12, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v27
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_or_b32_e32 v15, v15, v53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB29_2
-; GCN-NEXT: .LBB29_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v0, v54, v0
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_or_b32_e32 v4, v49, v4
-; GCN-NEXT: v_or_b32_e32 v5, v50, v5
-; GCN-NEXT: v_or_b32_e32 v6, v51, v6
-; GCN-NEXT: v_or_b32_e32 v7, v52, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v19, v9
-; GCN-NEXT: v_or_b32_e32 v10, v21, v10
-; GCN-NEXT: v_or_b32_e32 v11, v23, v11
-; GCN-NEXT: v_or_b32_e32 v12, v25, v12
-; GCN-NEXT: v_or_b32_e32 v13, v27, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_or_b32_e32 v15, v53, v15
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB29_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 3
-; VI-NEXT: v_add_u16_e32 v16, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: v_add_u16_e32 v16, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v16, v14
-; VI-NEXT: v_add_u16_e32 v16, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v16, v13
-; VI-NEXT: v_add_u16_e32 v16, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v16, v12
-; VI-NEXT: v_add_u16_e32 v16, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v16, v11
-; VI-NEXT: v_add_u16_e32 v16, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v16, v10
-; VI-NEXT: v_add_u16_e32 v16, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v16, v9
-; VI-NEXT: v_add_u16_e32 v16, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v16, v8
-; VI-NEXT: v_add_u16_e32 v16, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_add_u16_e32 v16, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v16, v6
-; VI-NEXT: v_add_u16_e32 v16, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v16, v5
-; VI-NEXT: v_add_u16_e32 v16, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v16, v4
-; VI-NEXT: v_add_u16_e32 v16, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v16, v3
-; VI-NEXT: v_add_u16_e32 v16, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v16, v2
-; VI-NEXT: v_add_u16_e32 v16, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v16, v1
-; VI-NEXT: v_add_u16_e32 v16, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB29_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB29_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB29_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB29_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB29_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <32 x half> @bitcast_v8i64_to_v32f16(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v34, v15
-; GCN-NEXT: v_mov_b32_e32 v33, v14
-; GCN-NEXT: v_mov_b32_e32 v36, v13
-; GCN-NEXT: v_mov_b32_e32 v35, v12
-; GCN-NEXT: v_mov_b32_e32 v38, v11
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v48, v9
-; GCN-NEXT: v_mov_b32_e32 v39, v8
-; GCN-NEXT: v_mov_b32_e32 v50, v7
-; GCN-NEXT: v_mov_b32_e32 v49, v6
-; GCN-NEXT: v_mov_b32_e32 v52, v5
-; GCN-NEXT: v_mov_b32_e32 v51, v4
-; GCN-NEXT: v_mov_b32_e32 v54, v3
-; GCN-NEXT: v_mov_b32_e32 v53, v2
-; GCN-NEXT: v_mov_b32_e32 v55, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v50
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v49
-; GCN-NEXT: s_waitcnt expcnt(5)
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v52
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v51
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v54
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v53
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: .LBB30_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB30_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v55, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v54, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v52, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v50, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v48, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v38, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v35
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v36, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v33
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v34, vcc
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32
-; GCN-NEXT: .LBB30_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB30_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB30_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB30_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB30_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v32f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB30_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB30_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <32 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <32 x half>
- br label %end
-
-end:
- %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x half> %phi
-}
-
-define <8 x i64> @bitcast_v32f16_to_v8i64(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43
-; GCN-NEXT: v_or_b32_e32 v0, v44, v0
-; GCN-NEXT: v_or_b32_e32 v1, v42, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_or_b32_e32 v3, v50, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v36, v6
-; GCN-NEXT: v_or_b32_e32 v7, v34, v7
-; GCN-NEXT: v_or_b32_e32 v8, v33, v8
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: v_or_b32_e32 v10, v31, v10
-; GCN-NEXT: v_or_b32_e32 v11, v21, v11
-; GCN-NEXT: v_or_b32_e32 v12, v19, v12
-; GCN-NEXT: v_or_b32_e32 v13, v18, v13
-; GCN-NEXT: v_or_b32_e32 v14, v17, v14
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB31_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB31_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v26, v24
-; GCN-NEXT: v_or_b32_e32 v10, v28, v27
-; GCN-NEXT: v_or_b32_e32 v11, v21, v29
-; GCN-NEXT: v_or_b32_e32 v12, v19, v25
-; GCN-NEXT: v_or_b32_e32 v13, v18, v23
-; GCN-NEXT: v_or_b32_e32 v14, v17, v22
-; GCN-NEXT: v_or_b32_e32 v15, v16, v20
-; GCN-NEXT: .LBB31_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB31_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 0x200
-; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v17
-; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v17
-; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v17
-; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v17
-; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v17
-; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v17
-; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v17
-; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v17
-; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v17
-; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v17
-; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v17
-; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v17
-; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v17
-; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v17
-; VI-NEXT: v_or_b32_e32 v0, v0, v16
-; VI-NEXT: .LBB31_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB31_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB31_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB31_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB31_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <32 x bfloat> @bitcast_v8i64_to_v32bf16(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v55, v15
-; GCN-NEXT: v_mov_b32_e32 v54, v14
-; GCN-NEXT: v_mov_b32_e32 v53, v13
-; GCN-NEXT: v_mov_b32_e32 v52, v12
-; GCN-NEXT: v_mov_b32_e32 v51, v11
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v49, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v39, v7
-; GCN-NEXT: v_mov_b32_e32 v38, v6
-; GCN-NEXT: v_mov_b32_e32 v37, v5
-; GCN-NEXT: v_mov_b32_e32 v36, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v3
-; GCN-NEXT: v_mov_b32_e32 v34, v2
-; GCN-NEXT: v_mov_b32_e32 v33, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB32_4
-; GCN-NEXT: .LBB32_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB32_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB32_2
-; GCN-NEXT: .LBB32_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v33, vcc
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34
-; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v35, vcc
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36
-; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v37, vcc
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38
-; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v39, vcc
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v49, vcc
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50
-; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v51, vcc
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52
-; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v53, vcc
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54
-; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v55, vcc
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB32_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
-; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
-; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
-; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
-; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
-; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
-; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
-; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
-; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
-; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
-; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT: .LBB32_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB32_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
-; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
-; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
-; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
-; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
-; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
-; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
-; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX9-NEXT: .LBB32_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB32_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB32_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <8 x i64> @bitcast_v32bf16_to_v8i64(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42
-; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16
-; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16
-; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16
-; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB33_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB33_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16
-; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16
-; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16
-; GCN-NEXT: .LBB33_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB33_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB33_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB33_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7
-; GFX9-NEXT: .LBB33_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32bf16_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB33_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19
-; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302
-; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23
-; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302
-; GFX11-NEXT: .LBB33_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <32 x bfloat> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x bfloat> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8i64_to_v64i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v23, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v50, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB34_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB34_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v23, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v50, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB34_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_or_b32_e32 v60, v1, v18
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17
-; GCN-NEXT: v_or_b32_e32 v17, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58
-; GCN-NEXT: v_or_b32_e32 v58, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v1, v2
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46
-; GCN-NEXT: v_or_b32_e32 v46, v1, v2
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59
-; GCN-NEXT: v_or_b32_e32 v44, v1, v2
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54
-; GCN-NEXT: v_or_b32_e32 v7, v2, v3
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47
-; GCN-NEXT: v_or_b32_e32 v54, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48
-; GCN-NEXT: v_or_b32_e32 v9, v4, v5
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v42
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55
-; GCN-NEXT: v_or_b32_e32 v48, v5, v6
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34
-; GCN-NEXT: v_or_b32_e32 v34, v6, v8
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v51
-; GCN-NEXT: v_or_b32_e32 v11, v8, v10
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28
-; GCN-NEXT: v_or_b32_e32 v13, v10, v12
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 24, v38
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35
-; GCN-NEXT: v_or_b32_e32 v28, v12, v14
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v23
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21
-; GCN-NEXT: v_or_b32_e32 v21, v14, v15
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v32
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29
-; GCN-NEXT: v_or_b32_e32 v29, v15, v16
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v32, v19, v16
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45
-; GCN-NEXT: v_or_b32_e32 v27, v61, v18
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41
-; GCN-NEXT: v_or_b32_e32 v40, v40, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57
-; GCN-NEXT: v_or_b32_e32 v56, v56, v22
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58
-; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46
-; GCN-NEXT: v_or_b32_e32 v49, v49, v47
-; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44
-; GCN-NEXT: v_or_b32_e32 v50, v50, v43
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_or_b32_e32 v36, v36, v55
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v37, v37, v53
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_or_b32_e32 v30, v30, v51
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48
-; GCN-NEXT: v_or_b32_e32 v31, v31, v39
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GCN-NEXT: v_or_b32_e32 v23, v23, v35
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v25, v25, v38
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v20, v20, v52
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_or_b32_e32 v26, v26, v42
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v33, v33, v45
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v24, v41, v24
-; GCN-NEXT: v_or_b32_e32 v32, v57, v32
-; GCN-NEXT: v_or_b32_e32 v17, v17, v27
-; GCN-NEXT: v_or_b32_e32 v27, v58, v40
-; GCN-NEXT: v_or_b32_e32 v35, v59, v56
-; GCN-NEXT: v_or_b32_e32 v38, v46, v49
-; GCN-NEXT: v_or_b32_e32 v39, v44, v50
-; GCN-NEXT: v_or_b32_e32 v7, v7, v36
-; GCN-NEXT: v_or_b32_e32 v36, v54, v37
-; GCN-NEXT: v_or_b32_e32 v9, v9, v30
-; GCN-NEXT: v_or_b32_e32 v30, v48, v31
-; GCN-NEXT: v_or_b32_e32 v23, v34, v23
-; GCN-NEXT: v_or_b32_e32 v11, v11, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v20
-; GCN-NEXT: v_or_b32_e32 v20, v28, v26
-; GCN-NEXT: v_or_b32_e32 v21, v21, v33
-; GCN-NEXT: v_or_b32_e32 v24, v29, v24
-; GCN-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v14, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v22, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8i64_to_v64i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_mov_b32_e32 v26, v22
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: .LBB34_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB34_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
-; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
-; VI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
-; VI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
-; VI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
-; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
-; VI-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
-; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
-; VI-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
-; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
-; VI-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: .LBB34_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8i64_to_v64i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_mov_b32_e32 v26, v23
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: .LBB34_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB34_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_co_u32_e32 v1, vcc, 3, v1
-; GFX9-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, 3, v3
-; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, 3, v5
-; GFX9-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v7, vcc, 3, v7
-; GFX9-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v8, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v9, vcc, 3, v9
-; GFX9-NEXT: v_addc_co_u32_e32 v10, vcc, 0, v10, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v11, vcc, 3, v11
-; GFX9-NEXT: v_addc_co_u32_e32 v12, vcc, 0, v12, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v13, vcc, 3, v13
-; GFX9-NEXT: v_addc_co_u32_e32 v14, vcc, 0, v14, vcc
-; GFX9-NEXT: v_add_co_u32_e32 v15, vcc, 3, v15
-; GFX9-NEXT: v_addc_co_u32_e32 v16, vcc, 0, v16, vcc
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB34_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8i64_to_v64i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB34_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB34_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_co_u32 v1, vcc_lo, v1, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, v3, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, v5, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v9, vcc_lo, v9, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v11, vcc_lo, v11, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v12, null, 0, v12, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v13, vcc_lo, v13, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v14, null, 0, v14, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v15, vcc_lo, v15, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_ci_u32_e64 v16, null, 0, v16, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v7, vcc_lo, v7, 3
-; GFX11-NEXT: v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB34_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v24, v96, v24
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v86, v85
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v24
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v84
-; GFX11-NEXT: v_or_b32_e32 v23, v83, v23
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v80
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v71
-; GFX11-NEXT: v_or_b32_e32 v22, v70, v22
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v24
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v23
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v25
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-NEXT: v_or_b32_e32 v21, v23, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v22, v25, v54
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v24
-; GFX11-NEXT: v_or_b32_e32 v23, v25, v49
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v48
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v22
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69
-; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v19, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v22
-; GFX11-NEXT: v_or_b32_e32 v18, v23, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v69
-; GFX11-NEXT: v_or_b32_e32 v67, v68, v67
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v22
-; GFX11-NEXT: v_or_b32_e32 v17, v23, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v20
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v21
-; GFX11-NEXT: s_clause 0x3
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <8 x i64> %a, splat (i64 3)
- %a2 = bitcast <8 x i64> %a1 to <64 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x i64> %a to <64 x i8>
- br label %end
-
-end:
- %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i8> %phi
-}
-
-define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i8_to_v8i64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v34, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v33, v8
-; GCN-NEXT: v_mov_b32_e32 v36, v6
-; GCN-NEXT: v_mov_b32_e32 v32, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_or_b32_e32 v0, v0, v42
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_or_b32_e32 v1, v1, v41
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v40
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_or_b32_e32 v3, v3, v55
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32
-; GCN-NEXT: v_or_b32_e32 v17, v33, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v28, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
-; GCN-NEXT: v_or_b32_e32 v27, v35, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v29, v36, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37
-; GCN-NEXT: v_or_b32_e32 v31, v38, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v9, v9, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v15, v15, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v25, v59, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v30, v57, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v32, v58, v8
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v8, v10
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v8, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v8, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v8, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v19, v46, v19
-; GCN-NEXT: v_or_b32_e32 v0, v0, v7
-; GCN-NEXT: v_or_b32_e32 v1, v1, v6
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_or_b32_e32 v4, v17, v21
-; GCN-NEXT: v_or_b32_e32 v5, v20, v22
-; GCN-NEXT: v_or_b32_e32 v6, v23, v24
-; GCN-NEXT: v_or_b32_e32 v7, v26, v28
-; GCN-NEXT: v_or_b32_e32 v8, v27, v25
-; GCN-NEXT: v_or_b32_e32 v9, v29, v30
-; GCN-NEXT: v_or_b32_e32 v10, v31, v32
-; GCN-NEXT: v_or_b32_e32 v11, v33, v34
-; GCN-NEXT: v_or_b32_e32 v12, v35, v12
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v16
-; GCN-NEXT: v_or_b32_e32 v15, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; kill: killed $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: .LBB35_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB35_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v42, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v41, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v40, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v55, v3
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7
-; GCN-NEXT: v_or_b32_e32 v6, v17, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v27, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v17, v44, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22
-; GCN-NEXT: v_or_b32_e32 v20, v45, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v56, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32
-; GCN-NEXT: v_or_b32_e32 v29, v43, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v19, v47, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_or_b32_e32 v16, v59, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_or_b32_e32 v18, v57, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_or_b32_e32 v22, v58, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19
-; GCN-NEXT: v_or_b32_e32 v30, v46, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v0, v31, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v10, v9
-; GCN-NEXT: v_or_b32_e32 v6, v12, v11
-; GCN-NEXT: v_or_b32_e32 v7, v14, v13
-; GCN-NEXT: v_or_b32_e32 v8, v16, v15
-; GCN-NEXT: v_or_b32_e32 v9, v18, v17
-; GCN-NEXT: v_or_b32_e32 v10, v22, v20
-; GCN-NEXT: v_or_b32_e32 v11, v24, v23
-; GCN-NEXT: v_or_b32_e32 v12, v26, v25
-; GCN-NEXT: v_or_b32_e32 v13, v28, v27
-; GCN-NEXT: v_or_b32_e32 v14, v21, v29
-; GCN-NEXT: v_or_b32_e32 v15, v30, v19
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15
-; GCN-NEXT: .LBB35_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i8_to_v8i64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: .LBB35_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB35_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v15, 0x300
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v51
-; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v10, 3, v24
-; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v30
-; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v60
-; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v42
-; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v53
-; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v52
-; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v50
-; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v63
-; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v48
-; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v44
-; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v39
-; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v16
-; VI-NEXT: v_add_u16_e32 v16, 3, v19
-; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: .LBB35_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i8_to_v8i64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: .LBB35_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB35_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v52
-; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v50
-; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: .LBB35_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i8_to_v8i64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v14, off, s32
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(25)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: s_waitcnt vmcnt(21)
-; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65
-; GFX11-NEXT: s_waitcnt vmcnt(19)
-; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67
-; GFX11-NEXT: s_waitcnt vmcnt(17)
-; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: s_waitcnt vmcnt(15)
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(13)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(11)
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB35_4
-; GFX11-NEXT: .LBB35_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB35_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v118
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v119
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v117
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v6, v113
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v114
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v116
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v101
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v102
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v103
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v112
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v100
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v11
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v87
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v96
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v97
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v98
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v99
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v81
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v82
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v83
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v84
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v85
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v27
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v29
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v65
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v66
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v67
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v17, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v18, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v19, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v20, v24, v25
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB35_2
-; GFX11-NEXT: .LBB35_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v118, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v119, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v2, v117, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v114, v3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v4, v115, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v116, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v113, v6
-; GFX11-NEXT: v_add_nc_u16 v7, v16, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v6, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v101, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v100, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v102, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v103, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v112, v10
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v6, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v10, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v87, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v96, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v97, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v98, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v99, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v81, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v82, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v83, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v84, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v85, v15
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_add_nc_u16 v11, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v11, v27, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v29, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v65, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v66, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v67, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v25, v24
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i8> %a, splat (i8 3)
- %a2 = bitcast <64 x i8> %a1 to <8 x i64>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i8> %a to <8 x i64>
- br label %end
-
-end:
- %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x i64> %phi
-}
-
-define <32 x i16> @bitcast_v8f64_to_v32i16(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v55, v15
-; GCN-NEXT: v_mov_b32_e32 v54, v14
-; GCN-NEXT: v_mov_b32_e32 v53, v13
-; GCN-NEXT: v_mov_b32_e32 v52, v12
-; GCN-NEXT: v_mov_b32_e32 v51, v11
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v49, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v38, v7
-; GCN-NEXT: v_mov_b32_e32 v37, v6
-; GCN-NEXT: v_mov_b32_e32 v36, v5
-; GCN-NEXT: v_mov_b32_e32 v35, v4
-; GCN-NEXT: v_mov_b32_e32 v34, v3
-; GCN-NEXT: v_mov_b32_e32 v33, v2
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16
-; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16
-; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16
-; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16
-; GCN-NEXT: v_alignbit_b32 v13, v38, v37, 16
-; GCN-NEXT: v_alignbit_b32 v9, v36, v35, 16
-; GCN-NEXT: v_alignbit_b32 v5, v34, v33, 16
-; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1
-; GCN-NEXT: .LBB36_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB36_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[33:34], v[33:34], 1.0
-; GCN-NEXT: v_add_f64 v[35:36], v[35:36], 1.0
-; GCN-NEXT: v_add_f64 v[37:38], v[37:38], 1.0
-; GCN-NEXT: v_add_f64 v[48:49], v[48:49], 1.0
-; GCN-NEXT: v_add_f64 v[50:51], v[50:51], 1.0
-; GCN-NEXT: v_add_f64 v[52:53], v[52:53], 1.0
-; GCN-NEXT: v_add_f64 v[54:55], v[54:55], 1.0
-; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16
-; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16
-; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16
-; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16
-; GCN-NEXT: v_alignbit_b32 v13, v38, v37, 16
-; GCN-NEXT: v_alignbit_b32 v9, v36, v35, 16
-; GCN-NEXT: v_alignbit_b32 v5, v34, v33, 16
-; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1
-; GCN-NEXT: .LBB36_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v2, v1
-; GCN-NEXT: v_mov_b32_e32 v4, v33
-; GCN-NEXT: v_mov_b32_e32 v6, v34
-; GCN-NEXT: v_mov_b32_e32 v8, v35
-; GCN-NEXT: v_mov_b32_e32 v10, v36
-; GCN-NEXT: v_mov_b32_e32 v12, v37
-; GCN-NEXT: v_mov_b32_e32 v14, v38
-; GCN-NEXT: v_mov_b32_e32 v16, v48
-; GCN-NEXT: v_mov_b32_e32 v18, v49
-; GCN-NEXT: v_mov_b32_e32 v20, v50
-; GCN-NEXT: v_mov_b32_e32 v22, v51
-; GCN-NEXT: v_mov_b32_e32 v24, v52
-; GCN-NEXT: v_mov_b32_e32 v26, v53
-; GCN-NEXT: v_mov_b32_e32 v28, v54
-; GCN-NEXT: v_mov_b32_e32 v30, v55
-; GCN-NEXT: v_mov_b32_e32 v1, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB36_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB36_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB36_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB36_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB36_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB36_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <8 x double> @bitcast_v32i16_to_v8f64(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v37, v12
-; GCN-NEXT: v_mov_b32_e32 v36, v10
-; GCN-NEXT: v_mov_b32_e32 v35, v8
-; GCN-NEXT: v_mov_b32_e32 v34, v6
-; GCN-NEXT: v_mov_b32_e32 v33, v4
-; GCN-NEXT: v_mov_b32_e32 v32, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB37_4
-; GCN-NEXT: .LBB37_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB37_3: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; GCN-NEXT: v_or_b32_e32 v0, v0, v54
-; GCN-NEXT: v_or_b32_e32 v1, v1, v55
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v2, v2, v39
-; GCN-NEXT: v_or_b32_e32 v3, v3, v48
-; GCN-NEXT: v_or_b32_e32 v4, v4, v49
-; GCN-NEXT: v_or_b32_e32 v5, v5, v50
-; GCN-NEXT: v_or_b32_e32 v6, v6, v51
-; GCN-NEXT: v_or_b32_e32 v7, v7, v52
-; GCN-NEXT: v_or_b32_e32 v8, v8, v17
-; GCN-NEXT: v_or_b32_e32 v9, v9, v19
-; GCN-NEXT: v_or_b32_e32 v10, v10, v21
-; GCN-NEXT: v_or_b32_e32 v11, v11, v23
-; GCN-NEXT: v_or_b32_e32 v12, v12, v25
-; GCN-NEXT: v_or_b32_e32 v13, v13, v27
-; GCN-NEXT: v_or_b32_e32 v14, v14, v29
-; GCN-NEXT: v_or_b32_e32 v15, v15, v53
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB37_2
-; GCN-NEXT: .LBB37_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v0, v54, v0
-; GCN-NEXT: v_or_b32_e32 v1, v55, v1
-; GCN-NEXT: v_or_b32_e32 v2, v39, v2
-; GCN-NEXT: v_or_b32_e32 v3, v48, v3
-; GCN-NEXT: v_or_b32_e32 v4, v49, v4
-; GCN-NEXT: v_or_b32_e32 v5, v50, v5
-; GCN-NEXT: v_or_b32_e32 v6, v51, v6
-; GCN-NEXT: v_or_b32_e32 v7, v52, v7
-; GCN-NEXT: v_or_b32_e32 v8, v17, v8
-; GCN-NEXT: v_or_b32_e32 v9, v19, v9
-; GCN-NEXT: v_or_b32_e32 v10, v21, v10
-; GCN-NEXT: v_or_b32_e32 v11, v23, v11
-; GCN-NEXT: v_or_b32_e32 v12, v25, v12
-; GCN-NEXT: v_or_b32_e32 v13, v27, v13
-; GCN-NEXT: v_or_b32_e32 v14, v29, v14
-; GCN-NEXT: v_or_b32_e32 v15, v53, v15
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB37_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 3
-; VI-NEXT: v_add_u16_e32 v16, 3, v15
-; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: v_add_u16_e32 v16, 3, v14
-; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v16, v14
-; VI-NEXT: v_add_u16_e32 v16, 3, v13
-; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v16, v13
-; VI-NEXT: v_add_u16_e32 v16, 3, v12
-; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v16, v12
-; VI-NEXT: v_add_u16_e32 v16, 3, v11
-; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v16, v11
-; VI-NEXT: v_add_u16_e32 v16, 3, v10
-; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v16, v10
-; VI-NEXT: v_add_u16_e32 v16, 3, v9
-; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v16, v9
-; VI-NEXT: v_add_u16_e32 v16, 3, v8
-; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v16, v8
-; VI-NEXT: v_add_u16_e32 v16, 3, v7
-; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v16, v7
-; VI-NEXT: v_add_u16_e32 v16, 3, v6
-; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v16, v6
-; VI-NEXT: v_add_u16_e32 v16, 3, v5
-; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v16, v5
-; VI-NEXT: v_add_u16_e32 v16, 3, v4
-; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v16, v4
-; VI-NEXT: v_add_u16_e32 v16, 3, v3
-; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, v16, v3
-; VI-NEXT: v_add_u16_e32 v16, 3, v2
-; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v16, v2
-; VI-NEXT: v_add_u16_e32 v16, 3, v1
-; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v1, v16, v1
-; VI-NEXT: v_add_u16_e32 v16, 3, v0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB37_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB37_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB37_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB37_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB37_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <32 x half> @bitcast_v8f64_to_v32f16(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v3
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v1
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: .LBB38_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB38_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v38, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v37, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v36, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v35, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v34, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v33, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v32, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v39, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v55, v55
-; GCN-NEXT: .LBB38_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v39
-; GCN-NEXT: v_mov_b32_e32 v1, v55
-; GCN-NEXT: v_mov_b32_e32 v2, v32
-; GCN-NEXT: v_mov_b32_e32 v3, v54
-; GCN-NEXT: v_mov_b32_e32 v4, v33
-; GCN-NEXT: v_mov_b32_e32 v5, v53
-; GCN-NEXT: v_mov_b32_e32 v6, v34
-; GCN-NEXT: v_mov_b32_e32 v7, v52
-; GCN-NEXT: v_mov_b32_e32 v8, v35
-; GCN-NEXT: v_mov_b32_e32 v9, v51
-; GCN-NEXT: v_mov_b32_e32 v10, v36
-; GCN-NEXT: v_mov_b32_e32 v11, v50
-; GCN-NEXT: v_mov_b32_e32 v12, v37
-; GCN-NEXT: v_mov_b32_e32 v13, v49
-; GCN-NEXT: v_mov_b32_e32 v14, v38
-; GCN-NEXT: v_mov_b32_e32 v15, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB38_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB38_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB38_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB38_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v32f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB38_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB38_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <32 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <32 x half>
- br label %end
-
-end:
- %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x half> %phi
-}
-
-define <8 x double> @bitcast_v32f16_to_v8f64(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43
-; GCN-NEXT: v_or_b32_e32 v0, v44, v0
-; GCN-NEXT: v_or_b32_e32 v1, v42, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v52, v2
-; GCN-NEXT: v_or_b32_e32 v3, v50, v3
-; GCN-NEXT: v_or_b32_e32 v4, v48, v4
-; GCN-NEXT: v_or_b32_e32 v5, v38, v5
-; GCN-NEXT: v_or_b32_e32 v6, v36, v6
-; GCN-NEXT: v_or_b32_e32 v7, v34, v7
-; GCN-NEXT: v_or_b32_e32 v8, v33, v8
-; GCN-NEXT: v_or_b32_e32 v9, v32, v9
-; GCN-NEXT: v_or_b32_e32 v10, v31, v10
-; GCN-NEXT: v_or_b32_e32 v11, v21, v11
-; GCN-NEXT: v_or_b32_e32 v12, v19, v12
-; GCN-NEXT: v_or_b32_e32 v13, v18, v13
-; GCN-NEXT: v_or_b32_e32 v14, v17, v14
-; GCN-NEXT: v_or_b32_e32 v15, v16, v15
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB39_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB39_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_or_b32_e32 v0, v1, v0
-; GCN-NEXT: v_or_b32_e32 v1, v3, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_or_b32_e32 v2, v3, v2
-; GCN-NEXT: v_or_b32_e32 v3, v5, v4
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v9, v8
-; GCN-NEXT: v_or_b32_e32 v6, v11, v10
-; GCN-NEXT: v_or_b32_e32 v7, v13, v12
-; GCN-NEXT: v_or_b32_e32 v8, v15, v14
-; GCN-NEXT: v_or_b32_e32 v9, v26, v24
-; GCN-NEXT: v_or_b32_e32 v10, v28, v27
-; GCN-NEXT: v_or_b32_e32 v11, v21, v29
-; GCN-NEXT: v_or_b32_e32 v12, v19, v25
-; GCN-NEXT: v_or_b32_e32 v13, v18, v23
-; GCN-NEXT: v_or_b32_e32 v14, v17, v22
-; GCN-NEXT: v_or_b32_e32 v15, v16, v20
-; GCN-NEXT: .LBB39_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB39_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 0x200
-; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v17
-; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v17
-; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v17
-; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v17
-; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v17
-; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v17
-; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v17
-; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v17
-; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v17
-; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v17
-; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v17
-; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v17
-; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
-; VI-NEXT: v_or_b32_e32 v2, v2, v17
-; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
-; VI-NEXT: v_or_b32_e32 v1, v1, v17
-; VI-NEXT: v_or_b32_e32 v0, v0, v16
-; VI-NEXT: .LBB39_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB39_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB39_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB39_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB39_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <32 x bfloat> @bitcast_v8f64_to_v32bf16(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v6
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v4
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v2
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v1
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: .LBB40_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB40_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v6
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v4
-; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v2
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v1
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v0
-; GCN-NEXT: .LBB40_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v55
-; GCN-NEXT: v_mov_b32_e32 v1, v54
-; GCN-NEXT: v_mov_b32_e32 v2, v53
-; GCN-NEXT: v_mov_b32_e32 v3, v52
-; GCN-NEXT: v_mov_b32_e32 v4, v51
-; GCN-NEXT: v_mov_b32_e32 v5, v50
-; GCN-NEXT: v_mov_b32_e32 v6, v49
-; GCN-NEXT: v_mov_b32_e32 v7, v48
-; GCN-NEXT: v_mov_b32_e32 v8, v39
-; GCN-NEXT: v_mov_b32_e32 v9, v38
-; GCN-NEXT: v_mov_b32_e32 v10, v37
-; GCN-NEXT: v_mov_b32_e32 v11, v36
-; GCN-NEXT: v_mov_b32_e32 v12, v35
-; GCN-NEXT: v_mov_b32_e32 v13, v34
-; GCN-NEXT: v_mov_b32_e32 v14, v33
-; GCN-NEXT: v_mov_b32_e32 v15, v32
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB40_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; VI-NEXT: .LBB40_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB40_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX9-NEXT: .LBB40_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB40_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-NEXT: .LBB40_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <8 x double> @bitcast_v32bf16_to_v8f64(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47
-; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46
-; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42
-; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16
-; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16
-; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16
-; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16
-; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16
-; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16
-; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16
-; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16
-; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16
-; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16
-; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: .LBB41_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB41_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16
-; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16
-; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16
-; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16
-; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16
-; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16
-; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16
-; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16
-; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16
-; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16
-; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16
-; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16
-; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16
-; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16
-; GCN-NEXT: .LBB41_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB41_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v15, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB41_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB41_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
-; GFX9-NEXT: s_mov_b32 s7, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7
-; GFX9-NEXT: .LBB41_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32bf16_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB41_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302
-; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8
-; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6
-; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1
-; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19
-; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302
-; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302
-; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
-; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
-; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2
-; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23
-; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0
-; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302
-; GFX11-NEXT: .LBB41_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <32 x bfloat> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x bfloat> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) {
-; GCN-LABEL: bitcast_v8f64_to_v64i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; kill: killed $vgpr18
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 8
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v24, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v22, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v21, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v30, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v27, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v28, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v34, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v32, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v48, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v37, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v39, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v40, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v52, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v42, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v54, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v55, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB42_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB42_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 8
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v24, v14, v13, 8
-; GCN-NEXT: v_alignbit_b32 v22, v12, v11, 24
-; GCN-NEXT: v_alignbit_b32 v21, v12, v11, 16
-; GCN-NEXT: v_alignbit_b32 v30, v12, v11, 8
-; GCN-NEXT: v_alignbit_b32 v27, v10, v9, 24
-; GCN-NEXT: v_alignbit_b32 v28, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v34, v10, v9, 8
-; GCN-NEXT: v_alignbit_b32 v32, v8, v7, 24
-; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16
-; GCN-NEXT: v_alignbit_b32 v48, v8, v7, 8
-; GCN-NEXT: v_alignbit_b32 v37, v6, v5, 24
-; GCN-NEXT: v_alignbit_b32 v39, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v40, v6, v5, 8
-; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 24
-; GCN-NEXT: v_alignbit_b32 v52, v4, v3, 16
-; GCN-NEXT: v_alignbit_b32 v42, v4, v3, 8
-; GCN-NEXT: v_alignbit_b32 v54, v2, v1, 24
-; GCN-NEXT: v_alignbit_b32 v55, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v16
-; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 24, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 24, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v41, 24, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v46, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2
-; GCN-NEXT: .LBB42_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v45
-; GCN-NEXT: v_or_b32_e32 v45, v1, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v17
-; GCN-NEXT: v_or_b32_e32 v18, v2, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v42
-; GCN-NEXT: v_or_b32_e32 v42, v3, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v63
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_or_b32_e32 v63, v4, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v40
-; GCN-NEXT: v_or_b32_e32 v40, v5, v1
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 24, v54
-; GCN-NEXT: v_and_b32_e32 v62, 0xff, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v60
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v61
-; GCN-NEXT: v_or_b32_e32 v14, v6, v1
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: v_and_b32_e32 v61, 0xff, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v48
-; GCN-NEXT: v_or_b32_e32 v15, v7, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v58, 0xff, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v59
-; GCN-NEXT: v_or_b32_e32 v16, v8, v3
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v52, 0xff, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v34
-; GCN-NEXT: v_or_b32_e32 v17, v9, v4
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v56
-; GCN-NEXT: v_or_b32_e32 v34, v10, v5
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v30
-; GCN-NEXT: v_or_b32_e32 v30, v11, v6
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v44
-; GCN-NEXT: v_or_b32_e32 v11, v12, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 28, v0
-; GCN-NEXT: v_and_b32_e32 v41, 0xff, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v24
-; GCN-NEXT: v_or_b32_e32 v13, v13, v8
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v50
-; GCN-NEXT: v_or_b32_e32 v24, v19, v9
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0
-; GCN-NEXT: v_and_b32_e32 v50, 0xff, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v22
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v10
-; GCN-NEXT: v_or_b32_e32 v21, v20, v10
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v35
-; GCN-NEXT: v_or_b32_e32 v23, v23, v12
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 44, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v44, 0xff, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v55
-; GCN-NEXT: v_or_b32_e32 v36, v54, v19
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 48, v0
-; GCN-NEXT: v_and_b32_e32 v54, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v62
-; GCN-NEXT: v_or_b32_e32 v31, v60, v20
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v28
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v61
-; GCN-NEXT: v_or_b32_e32 v56, v25, v28
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 56, v0
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v58
-; GCN-NEXT: v_or_b32_e32 v57, v57, v25
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 60, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v42, 0xffff, v42
-; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v63
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v40
-; GCN-NEXT: v_or_b32_e32 v37, v37, v52
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v39, v39, v47
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_or_b32_e32 v32, v32, v46
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v33, v33, v43
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v27, v27, v41
-; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GCN-NEXT: v_or_b32_e32 v48, v48, v53
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v49, v49, v50
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_or_b32_e32 v22, v22, v51
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v35, v35, v44
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v29, v29, v54
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v38, v38, v55
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_or_b32_e32 v26, v59, v26
-; GCN-NEXT: v_or_b32_e32 v36, v45, v36
-; GCN-NEXT: v_or_b32_e32 v18, v18, v31
-; GCN-NEXT: v_or_b32_e32 v31, v42, v56
-; GCN-NEXT: v_or_b32_e32 v50, v58, v57
-; GCN-NEXT: v_or_b32_e32 v37, v40, v37
-; GCN-NEXT: v_or_b32_e32 v14, v14, v39
-; GCN-NEXT: v_or_b32_e32 v15, v15, v32
-; GCN-NEXT: v_or_b32_e32 v16, v16, v33
-; GCN-NEXT: v_or_b32_e32 v17, v17, v27
-; GCN-NEXT: v_or_b32_e32 v27, v34, v48
-; GCN-NEXT: v_or_b32_e32 v30, v30, v49
-; GCN-NEXT: v_or_b32_e32 v11, v11, v22
-; GCN-NEXT: v_or_b32_e32 v13, v13, v35
-; GCN-NEXT: v_or_b32_e32 v22, v24, v29
-; GCN-NEXT: v_or_b32_e32 v21, v21, v38
-; GCN-NEXT: v_or_b32_e32 v23, v23, v26
-; GCN-NEXT: buffer_store_dword v36, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v18, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v31, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v50, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v37, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v5, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v15, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v7, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v8, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v27, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v10, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v11, v12, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v20, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v8f64_to_v64i8:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_mov_b32_e32 v26, v22
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: .LBB42_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB42_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; VI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; VI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; VI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; VI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; VI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; VI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; VI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: .LBB42_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22
-; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24
-; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v8f64_to_v64i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_mov_b32_e32 v26, v23
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: .LBB42_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB42_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GFX9-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GFX9-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GFX9-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX9-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX9-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX9-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX9-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB42_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v8f64_to_v64i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB42_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB42_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB42_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v24, v96, v24
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v86, v85
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v24
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v84
-; GFX11-NEXT: v_or_b32_e32 v23, v83, v23
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v80
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v71
-; GFX11-NEXT: v_or_b32_e32 v22, v70, v22
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v24
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v23
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v25
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-NEXT: v_or_b32_e32 v21, v23, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v22, v25, v54
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v24
-; GFX11-NEXT: v_or_b32_e32 v23, v25, v49
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v48
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v22
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69
-; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v19, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v22
-; GFX11-NEXT: v_or_b32_e32 v18, v23, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v69
-; GFX11-NEXT: v_or_b32_e32 v67, v68, v67
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v22
-; GFX11-NEXT: v_or_b32_e32 v17, v23, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v20
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v21
-; GFX11-NEXT: s_clause 0x3
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <8 x double> %a, splat (double 1.000000e+00)
- %a2 = bitcast <8 x double> %a1 to <64 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <8 x double> %a to <64 x i8>
- br label %end
-
-end:
- %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i8> %phi
-}
-
-define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i8_to_v8f64:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: v_mov_b32_e32 v38, v14
-; GCN-NEXT: v_mov_b32_e32 v34, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v33, v8
-; GCN-NEXT: v_mov_b32_e32 v36, v6
-; GCN-NEXT: v_mov_b32_e32 v32, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v2
-; GCN-NEXT: v_mov_b32_e32 v31, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44
-; GCN-NEXT: s_waitcnt vmcnt(10)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46
-; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GCN-NEXT: v_or_b32_e32 v0, v0, v42
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_or_b32_e32 v1, v1, v41
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_or_b32_e32 v2, v2, v40
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GCN-NEXT: v_or_b32_e32 v3, v3, v55
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54
-; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52
-; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32
-; GCN-NEXT: v_or_b32_e32 v17, v33, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v20, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v24, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: v_or_b32_e32 v26, v28, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
-; GCN-NEXT: v_or_b32_e32 v27, v35, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_or_b32_e32 v29, v36, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37
-; GCN-NEXT: v_or_b32_e32 v31, v38, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_or_b32_e32 v9, v9, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_or_b32_e32 v15, v15, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_or_b32_e32 v18, v18, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_or_b32_e32 v25, v59, v25
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v30, v57, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v32, v58, v8
-; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v34, v8, v10
-; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v8, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v8, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v8, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_or_b32_e32 v19, v46, v19
-; GCN-NEXT: v_or_b32_e32 v0, v0, v7
-; GCN-NEXT: v_or_b32_e32 v1, v1, v6
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_or_b32_e32 v4, v17, v21
-; GCN-NEXT: v_or_b32_e32 v5, v20, v22
-; GCN-NEXT: v_or_b32_e32 v6, v23, v24
-; GCN-NEXT: v_or_b32_e32 v7, v26, v28
-; GCN-NEXT: v_or_b32_e32 v8, v27, v25
-; GCN-NEXT: v_or_b32_e32 v9, v29, v30
-; GCN-NEXT: v_or_b32_e32 v10, v31, v32
-; GCN-NEXT: v_or_b32_e32 v11, v33, v34
-; GCN-NEXT: v_or_b32_e32 v12, v35, v12
-; GCN-NEXT: v_or_b32_e32 v13, v13, v14
-; GCN-NEXT: v_or_b32_e32 v14, v15, v16
-; GCN-NEXT: v_or_b32_e32 v15, v18, v19
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; kill: killed $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; kill: killed $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; kill: killed $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: .LBB43_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB43_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GCN-NEXT: v_or_b32_e32 v0, v42, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v41, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_or_b32_e32 v2, v40, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v55, v3
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7
-; GCN-NEXT: v_or_b32_e32 v6, v17, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v11, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
-; GCN-NEXT: v_or_b32_e32 v13, v29, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
-; GCN-NEXT: v_or_b32_e32 v15, v27, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18
-; GCN-NEXT: v_or_b32_e32 v17, v44, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22
-; GCN-NEXT: v_or_b32_e32 v20, v45, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
-; GCN-NEXT: v_or_b32_e32 v23, v56, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32
-; GCN-NEXT: v_or_b32_e32 v29, v43, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_or_b32_e32 v19, v47, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v36
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v32, v5
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v32, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v7, v32, v7
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v32, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15
-; GCN-NEXT: v_or_b32_e32 v16, v59, v16
-; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17
-; GCN-NEXT: v_or_b32_e32 v18, v57, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: v_or_b32_e32 v22, v58, v22
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v32, v26
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19
-; GCN-NEXT: v_or_b32_e32 v30, v46, v30
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_or_b32_e32 v0, v31, v0
-; GCN-NEXT: v_or_b32_e32 v1, v8, v1
-; GCN-NEXT: v_or_b32_e32 v2, v5, v2
-; GCN-NEXT: v_or_b32_e32 v3, v4, v3
-; GCN-NEXT: v_or_b32_e32 v4, v7, v6
-; GCN-NEXT: v_or_b32_e32 v5, v10, v9
-; GCN-NEXT: v_or_b32_e32 v6, v12, v11
-; GCN-NEXT: v_or_b32_e32 v7, v14, v13
-; GCN-NEXT: v_or_b32_e32 v8, v16, v15
-; GCN-NEXT: v_or_b32_e32 v9, v18, v17
-; GCN-NEXT: v_or_b32_e32 v10, v22, v20
-; GCN-NEXT: v_or_b32_e32 v11, v24, v23
-; GCN-NEXT: v_or_b32_e32 v12, v26, v25
-; GCN-NEXT: v_or_b32_e32 v13, v28, v27
-; GCN-NEXT: v_or_b32_e32 v14, v21, v29
-; GCN-NEXT: v_or_b32_e32 v15, v30, v19
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15
-; GCN-NEXT: .LBB43_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i8_to_v8f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB43_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: .LBB43_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB43_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v15, 0x300
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u16_e32 v9, 3, v51
-; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v10, 3, v24
-; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v30
-; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v60
-; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v47
-; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v42
-; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v17, 3, v17
-; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 0x300, v0
-; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v0, v0, v1
-; VI-NEXT: v_or_b32_e32 v1, v2, v3
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 3, v53
-; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v2, 0x300, v2
-; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v2, v2, v3
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v4
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v4, 0x300, v4
-; VI-NEXT: v_or_b32_e32 v4, v4, v5
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v6
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v7
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v7, 0x300, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v8
-; VI-NEXT: v_add_u16_e32 v8, 3, v52
-; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v8, 0x300, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v9
-; VI-NEXT: v_add_u16_e32 v9, 3, v50
-; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v9, 0x300, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v10
-; VI-NEXT: v_add_u16_e32 v10, 3, v49
-; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v10, 0x300, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v11
-; VI-NEXT: v_add_u16_e32 v11, 3, v63
-; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v11, 0x300, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v12
-; VI-NEXT: v_add_u16_e32 v12, 3, v48
-; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v12, 0x300, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v13
-; VI-NEXT: v_add_u16_e32 v13, 3, v44
-; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v13, 0x300, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v14
-; VI-NEXT: v_add_u16_e32 v14, 3, v39
-; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v16
-; VI-NEXT: v_add_u16_e32 v16, 3, v19
-; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v15, v16, v15
-; VI-NEXT: .LBB43_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i8_to_v8f64:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124
-; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(25)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB43_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: .LBB43_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB43_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: s_movk_i32 s6, 0x300
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v24
-; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX9-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2
-; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_e32 v2, v2, v3
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3
-; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v4, 3, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4
-; GFX9-NEXT: v_or_b32_e32 v4, v4, v5
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v5, 3, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5
-; GFX9-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7
-; GFX9-NEXT: v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT: v_add_u16_e32 v8, 3, v52
-; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8
-; GFX9-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT: v_add_u16_e32 v9, 3, v50
-; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9
-; GFX9-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT: v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10
-; GFX9-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT: v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11
-; GFX9-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT: v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12
-; GFX9-NEXT: v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT: v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13
-; GFX9-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT: v_add_u16_e32 v14, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14
-; GFX9-NEXT: v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT: v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15
-; GFX9-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT: .LBB43_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i8_to_v8f64:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
-; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v14, off, s32
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(29)
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4
-; GFX11-NEXT: s_waitcnt vmcnt(27)
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6
-; GFX11-NEXT: s_waitcnt vmcnt(25)
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: s_waitcnt vmcnt(21)
-; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65
-; GFX11-NEXT: s_waitcnt vmcnt(19)
-; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67
-; GFX11-NEXT: s_waitcnt vmcnt(17)
-; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: s_waitcnt vmcnt(15)
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-NEXT: s_waitcnt vmcnt(13)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130
-; GFX11-NEXT: s_waitcnt vmcnt(11)
-; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB43_4
-; GFX11-NEXT: .LBB43_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB43_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v118
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v119
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v117
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v6, v113
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v114
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v116
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v101
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v102
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v103
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v112
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v100
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v5
-; GFX11-NEXT: v_or_b32_e32 v3, v6, v11
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v87
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v96
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v97
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v98
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v99
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v81
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v82
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v83
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v84
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v85
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v27
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v29
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v65
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v66
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v67
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX11-NEXT: v_or_b32_e32 v17, v18, v19
-; GFX11-NEXT: v_or_b32_e32 v18, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v19, v22, v23
-; GFX11-NEXT: v_or_b32_e32 v20, v24, v25
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB43_2
-; GFX11-NEXT: .LBB43_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v31, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v32, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v4, v35, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v5, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v6, v37, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v118, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v119, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_or_b32_e32 v2, v117, v2
-; GFX11-NEXT: v_or_b32_e32 v3, v114, v3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_or_b32_e32 v4, v115, v4
-; GFX11-NEXT: v_or_b32_e32 v5, v116, v5
-; GFX11-NEXT: v_or_b32_e32 v6, v113, v6
-; GFX11-NEXT: v_add_nc_u16 v7, v16, 3
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v6, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v18, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v20, 3
-; GFX11-NEXT: v_add_nc_u16 v10, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_or_b32_e32 v7, v101, v7
-; GFX11-NEXT: v_or_b32_e32 v6, v100, v6
-; GFX11-NEXT: v_or_b32_e32 v8, v102, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v103, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v112, v10
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v3, v5, v6
-; GFX11-NEXT: v_or_b32_e32 v4, v7, v8
-; GFX11-NEXT: v_or_b32_e32 v5, v9, v10
-; GFX11-NEXT: v_add_nc_u16 v6, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v7, v26, 3
-; GFX11-NEXT: v_add_nc_u16 v8, v28, 3
-; GFX11-NEXT: v_add_nc_u16 v9, v30, 3
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v10, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v11, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v71, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v68, 3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v87, v6
-; GFX11-NEXT: v_or_b32_e32 v7, v96, v7
-; GFX11-NEXT: v_or_b32_e32 v8, v97, v8
-; GFX11-NEXT: v_or_b32_e32 v9, v98, v9
-; GFX11-NEXT: v_or_b32_e32 v10, v99, v10
-; GFX11-NEXT: v_or_b32_e32 v11, v81, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v82, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v83, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v84, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v85, v15
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v7
-; GFX11-NEXT: v_or_b32_e32 v7, v8, v9
-; GFX11-NEXT: v_or_b32_e32 v8, v10, v11
-; GFX11-NEXT: v_or_b32_e32 v9, v12, v13
-; GFX11-NEXT: v_or_b32_e32 v10, v14, v15
-; GFX11-NEXT: v_add_nc_u16 v11, v64, 3
-; GFX11-NEXT: v_add_nc_u16 v12, v55, 3
-; GFX11-NEXT: v_add_nc_u16 v13, v54, 3
-; GFX11-NEXT: v_add_nc_u16 v14, v53, 3
-; GFX11-NEXT: v_add_nc_u16 v15, v52, 3
-; GFX11-NEXT: v_add_nc_u16 v16, v51, 3
-; GFX11-NEXT: v_add_nc_u16 v18, v50, 3
-; GFX11-NEXT: v_add_nc_u16 v20, v49, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v48, 3
-; GFX11-NEXT: v_add_nc_u16 v24, v39, 3
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-NEXT: v_or_b32_e32 v11, v27, v11
-; GFX11-NEXT: v_or_b32_e32 v12, v29, v12
-; GFX11-NEXT: v_or_b32_e32 v13, v65, v13
-; GFX11-NEXT: v_or_b32_e32 v14, v66, v14
-; GFX11-NEXT: v_or_b32_e32 v15, v67, v15
-; GFX11-NEXT: v_or_b32_e32 v16, v17, v16
-; GFX11-NEXT: v_or_b32_e32 v17, v19, v18
-; GFX11-NEXT: v_or_b32_e32 v18, v21, v20
-; GFX11-NEXT: v_or_b32_e32 v19, v23, v22
-; GFX11-NEXT: v_or_b32_e32 v20, v25, v24
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v12
-; GFX11-NEXT: v_or_b32_e32 v12, v13, v14
-; GFX11-NEXT: v_or_b32_e32 v13, v15, v16
-; GFX11-NEXT: v_or_b32_e32 v14, v17, v18
-; GFX11-NEXT: v_or_b32_e32 v15, v19, v20
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i8> %a, splat (i8 3)
- %a2 = bitcast <64 x i8> %a1 to <8 x double>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i8> %a to <8 x double>
- br label %end
-
-end:
- %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <8 x double> %phi
-}
-
-define <32 x half> @bitcast_v32i16_to_v32f16(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_mov_b32_e32 v62, v30
-; GCN-NEXT: v_mov_b32_e32 v61, v29
-; GCN-NEXT: v_mov_b32_e32 v60, v28
-; GCN-NEXT: v_mov_b32_e32 v59, v27
-; GCN-NEXT: v_mov_b32_e32 v58, v26
-; GCN-NEXT: v_mov_b32_e32 v57, v25
-; GCN-NEXT: v_mov_b32_e32 v56, v24
-; GCN-NEXT: v_mov_b32_e32 v47, v23
-; GCN-NEXT: v_mov_b32_e32 v46, v22
-; GCN-NEXT: v_mov_b32_e32 v45, v21
-; GCN-NEXT: v_mov_b32_e32 v44, v20
-; GCN-NEXT: v_mov_b32_e32 v43, v19
-; GCN-NEXT: v_mov_b32_e32 v42, v18
-; GCN-NEXT: v_mov_b32_e32 v41, v17
-; GCN-NEXT: v_mov_b32_e32 v40, v16
-; GCN-NEXT: v_mov_b32_e32 v55, v15
-; GCN-NEXT: v_mov_b32_e32 v54, v14
-; GCN-NEXT: v_mov_b32_e32 v53, v13
-; GCN-NEXT: v_mov_b32_e32 v52, v12
-; GCN-NEXT: v_mov_b32_e32 v51, v11
-; GCN-NEXT: v_mov_b32_e32 v50, v10
-; GCN-NEXT: v_mov_b32_e32 v49, v9
-; GCN-NEXT: v_mov_b32_e32 v48, v8
-; GCN-NEXT: v_mov_b32_e32 v39, v7
-; GCN-NEXT: v_mov_b32_e32 v38, v6
-; GCN-NEXT: v_mov_b32_e32 v37, v5
-; GCN-NEXT: v_mov_b32_e32 v36, v4
-; GCN-NEXT: v_mov_b32_e32 v35, v3
-; GCN-NEXT: v_mov_b32_e32 v34, v2
-; GCN-NEXT: v_mov_b32_e32 v33, v1
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v62
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB44_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB44_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v60
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v59
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v58
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v57
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v46
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v45
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v44
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v43
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v42
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v41
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v55
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: .LBB44_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB44_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 3
-; VI-NEXT: v_add_u16_sdwa v19, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v19
-; VI-NEXT: v_add_u16_sdwa v19, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v19
-; VI-NEXT: v_add_u16_sdwa v19, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v19
-; VI-NEXT: v_add_u16_sdwa v19, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v19
-; VI-NEXT: v_add_u16_sdwa v19, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v19
-; VI-NEXT: v_add_u16_sdwa v19, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 3, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v19
-; VI-NEXT: v_add_u16_sdwa v19, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 3, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v19
-; VI-NEXT: v_add_u16_sdwa v19, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v19
-; VI-NEXT: v_add_u16_sdwa v19, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v19
-; VI-NEXT: v_add_u16_sdwa v19, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v19
-; VI-NEXT: v_add_u16_sdwa v19, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v19
-; VI-NEXT: v_add_u16_sdwa v19, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_add_u16_sdwa v17, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v19
-; VI-NEXT: v_add_u16_sdwa v19, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v16, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v3, v3, v16
-; VI-NEXT: v_or_b32_e32 v2, v2, v19
-; VI-NEXT: v_or_b32_e32 v1, v1, v18
-; VI-NEXT: v_or_b32_e32 v0, v0, v17
-; VI-NEXT: .LBB44_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB44_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB44_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v32f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB44_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB44_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <32 x half>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <32 x half>
- br label %end
-
-end:
- %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x half> %phi
-}
-
-define <32 x i16> @bitcast_v32f16_to_v32i16(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v32
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB45_2
-; GCN-NEXT: ; %bb.1: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v31
-; GCN-NEXT: v_or_b32_e32 v30, v30, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v27
-; GCN-NEXT: v_or_b32_e32 v26, v26, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v23
-; GCN-NEXT: v_or_b32_e32 v22, v22, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v19
-; GCN-NEXT: v_or_b32_e32 v18, v18, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GCN-NEXT: v_or_b32_e32 v14, v14, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v11
-; GCN-NEXT: v_or_b32_e32 v10, v10, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v7
-; GCN-NEXT: v_or_b32_e32 v6, v6, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v32
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: v_or_b32_e32 v0, v0, v1
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_or_b32_e32 v8, v8, v9
-; GCN-NEXT: v_or_b32_e32 v12, v12, v13
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_or_b32_e32 v20, v20, v21
-; GCN-NEXT: v_or_b32_e32 v24, v24, v25
-; GCN-NEXT: v_or_b32_e32 v28, v28, v29
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v21, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v25, 16
-; GCN-NEXT: v_alignbit_b32 v29, v30, v29, 16
-; GCN-NEXT: .LBB45_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB45_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 0x200
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v15
-; VI-NEXT: v_add_f16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v19, v15
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v14
-; VI-NEXT: v_add_f16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v19, v14
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v13
-; VI-NEXT: v_add_f16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v19, v13
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v12
-; VI-NEXT: v_add_f16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v19, v12
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v11
-; VI-NEXT: v_add_f16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v19, v11
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v10
-; VI-NEXT: v_add_f16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v19, v10
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v9
-; VI-NEXT: v_add_f16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v19, v9
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v19, v8
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v19, v7
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v19, v6
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v19, v5
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v19, v4
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v17, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_or_b32_e32 v2, v19, v2
-; VI-NEXT: v_or_b32_e32 v1, v18, v1
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB45_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB45_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB45_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB45_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB45_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <32 x bfloat> @bitcast_v32i16_to_v32bf16(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v55, v30
-; GCN-NEXT: v_mov_b32_e32 v54, v28
-; GCN-NEXT: v_mov_b32_e32 v53, v26
-; GCN-NEXT: v_mov_b32_e32 v52, v24
-; GCN-NEXT: v_mov_b32_e32 v51, v22
-; GCN-NEXT: v_mov_b32_e32 v50, v20
-; GCN-NEXT: v_mov_b32_e32 v49, v18
-; GCN-NEXT: v_mov_b32_e32 v48, v16
-; GCN-NEXT: v_mov_b32_e32 v39, v14
-; GCN-NEXT: v_mov_b32_e32 v38, v12
-; GCN-NEXT: v_mov_b32_e32 v37, v10
-; GCN-NEXT: v_mov_b32_e32 v36, v8
-; GCN-NEXT: v_mov_b32_e32 v35, v6
-; GCN-NEXT: v_mov_b32_e32 v34, v4
-; GCN-NEXT: v_mov_b32_e32 v33, v2
-; GCN-NEXT: v_mov_b32_e32 v32, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_3
-; GCN-NEXT: ; %bb.1: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB46_4
-; GCN-NEXT: .LBB46_2: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-; GCN-NEXT: .LBB46_3: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB46_2
-; GCN-NEXT: .LBB46_4: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v54
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v53
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v52
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v51
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v49
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v48
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v39
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v37
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v36
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v35
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v34
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v33
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v32
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GCN-NEXT: v_or_b32_e32 v0, v31, v0
-; GCN-NEXT: v_or_b32_e32 v2, v29, v2
-; GCN-NEXT: v_or_b32_e32 v4, v27, v4
-; GCN-NEXT: v_or_b32_e32 v6, v25, v6
-; GCN-NEXT: v_or_b32_e32 v8, v23, v8
-; GCN-NEXT: v_or_b32_e32 v10, v21, v10
-; GCN-NEXT: v_or_b32_e32 v12, v19, v12
-; GCN-NEXT: v_or_b32_e32 v14, v17, v14
-; GCN-NEXT: v_or_b32_e32 v15, v15, v16
-; GCN-NEXT: v_or_b32_e32 v13, v13, v18
-; GCN-NEXT: v_or_b32_e32 v11, v11, v20
-; GCN-NEXT: v_or_b32_e32 v9, v9, v22
-; GCN-NEXT: v_or_b32_e32 v7, v7, v24
-; GCN-NEXT: v_or_b32_e32 v5, v5, v26
-; GCN-NEXT: v_or_b32_e32 v3, v3, v28
-; GCN-NEXT: v_or_b32_e32 v1, v1, v30
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v0
-; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v10
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v14
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB46_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v16, 3
-; VI-NEXT: v_add_u16_sdwa v19, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v15, 3, v15
-; VI-NEXT: v_or_b32_e32 v15, v15, v19
-; VI-NEXT: v_add_u16_sdwa v19, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v14, 3, v14
-; VI-NEXT: v_or_b32_e32 v14, v14, v19
-; VI-NEXT: v_add_u16_sdwa v19, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v13, 3, v13
-; VI-NEXT: v_or_b32_e32 v13, v13, v19
-; VI-NEXT: v_add_u16_sdwa v19, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v12, 3, v12
-; VI-NEXT: v_or_b32_e32 v12, v12, v19
-; VI-NEXT: v_add_u16_sdwa v19, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v11, 3, v11
-; VI-NEXT: v_or_b32_e32 v11, v11, v19
-; VI-NEXT: v_add_u16_sdwa v19, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v10, 3, v10
-; VI-NEXT: v_or_b32_e32 v10, v10, v19
-; VI-NEXT: v_add_u16_sdwa v19, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v9, 3, v9
-; VI-NEXT: v_or_b32_e32 v9, v9, v19
-; VI-NEXT: v_add_u16_sdwa v19, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v8, 3, v8
-; VI-NEXT: v_or_b32_e32 v8, v8, v19
-; VI-NEXT: v_add_u16_sdwa v19, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v7, 3, v7
-; VI-NEXT: v_or_b32_e32 v7, v7, v19
-; VI-NEXT: v_add_u16_sdwa v19, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v6, 3, v6
-; VI-NEXT: v_or_b32_e32 v6, v6, v19
-; VI-NEXT: v_add_u16_sdwa v19, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v5, 3, v5
-; VI-NEXT: v_or_b32_e32 v5, v5, v19
-; VI-NEXT: v_add_u16_sdwa v19, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v4, 3, v4
-; VI-NEXT: v_add_u16_sdwa v17, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v18, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v4, v19
-; VI-NEXT: v_add_u16_sdwa v19, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v16, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_add_u16_e32 v2, 3, v2
-; VI-NEXT: v_add_u16_e32 v1, 3, v1
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_e32 v3, v3, v16
-; VI-NEXT: v_or_b32_e32 v2, v2, v19
-; VI-NEXT: v_or_b32_e32 v1, v1, v18
-; VI-NEXT: v_or_b32_e32 v0, v0, v17
-; VI-NEXT: .LBB46_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB46_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB46_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB46_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: .LBB46_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <32 x i16> @bitcast_v32bf16_to_v32i16(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:4
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v28
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v55
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v31
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v63
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: .LBB47_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB47_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v58
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v56
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v46
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v42
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v32
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15
-; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v19
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v20
-; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v22
-; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v24
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff0000, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v26
-; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v28
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GCN-NEXT: v_alignbit_b32 v0, v30, v0, 16
-; GCN-NEXT: v_alignbit_b32 v4, v36, v2, 16
-; GCN-NEXT: v_alignbit_b32 v8, v37, v32, 16
-; GCN-NEXT: v_alignbit_b32 v12, v38, v5, 16
-; GCN-NEXT: v_alignbit_b32 v16, v39, v33, 16
-; GCN-NEXT: v_alignbit_b32 v20, v48, v9, 16
-; GCN-NEXT: v_alignbit_b32 v24, v49, v10, 16
-; GCN-NEXT: v_alignbit_b32 v28, v50, v13, 16
-; GCN-NEXT: v_alignbit_b32 v30, v31, v14, 16
-; GCN-NEXT: v_alignbit_b32 v26, v27, v17, 16
-; GCN-NEXT: v_alignbit_b32 v22, v23, v18, 16
-; GCN-NEXT: v_alignbit_b32 v18, v19, v21, 16
-; GCN-NEXT: v_alignbit_b32 v14, v15, v34, 16
-; GCN-NEXT: v_alignbit_b32 v10, v11, v25, 16
-; GCN-NEXT: v_alignbit_b32 v6, v7, v35, 16
-; GCN-NEXT: v_alignbit_b32 v2, v3, v29, 16
-; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
-; GCN-NEXT: v_alignbit_b32 v5, v6, v41, 16
-; GCN-NEXT: v_alignbit_b32 v9, v10, v40, 16
-; GCN-NEXT: v_alignbit_b32 v13, v14, v55, 16
-; GCN-NEXT: v_alignbit_b32 v17, v18, v54, 16
-; GCN-NEXT: v_alignbit_b32 v21, v22, v53, 16
-; GCN-NEXT: v_alignbit_b32 v25, v26, v52, 16
-; GCN-NEXT: v_alignbit_b32 v29, v30, v51, 16
-; GCN-NEXT: .LBB47_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB47_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_bfe_u32 v18, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17
-; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc
-; VI-NEXT: v_bfe_u32 v18, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v1
-; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18
-; VI-NEXT: v_or_b32_e32 v19, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_bfe_u32 v19, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v18
-; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc
-; VI-NEXT: v_bfe_u32 v19, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v2
-; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19
-; VI-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_bfe_u32 v20, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v19
-; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc
-; VI-NEXT: v_bfe_u32 v20, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v3
-; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20
-; VI-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v4
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_bfe_u32 v21, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v20
-; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc
-; VI-NEXT: v_bfe_u32 v21, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v4
-; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21
-; VI-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_bfe_u32 v22, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v21
-; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc
-; VI-NEXT: v_bfe_u32 v22, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v5
-; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22
-; VI-NEXT: v_or_b32_e32 v23, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v6
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_bfe_u32 v23, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v22
-; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v24, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc
-; VI-NEXT: v_bfe_u32 v23, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v6
-; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23
-; VI-NEXT: v_or_b32_e32 v24, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v7
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_bfe_u32 v24, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v23
-; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v25, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc
-; VI-NEXT: v_bfe_u32 v24, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v7
-; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24
-; VI-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v8
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_bfe_u32 v25, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v24
-; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v26, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc
-; VI-NEXT: v_bfe_u32 v25, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v8
-; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25
-; VI-NEXT: v_or_b32_e32 v26, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v9
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_bfe_u32 v26, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v25
-; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v27, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc
-; VI-NEXT: v_bfe_u32 v26, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v9
-; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26
-; VI-NEXT: v_or_b32_e32 v27, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v10
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_bfe_u32 v27, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v26
-; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v28, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc
-; VI-NEXT: v_bfe_u32 v27, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v10
-; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27
-; VI-NEXT: v_or_b32_e32 v28, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v11
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_bfe_u32 v28, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v27
-; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v29, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc
-; VI-NEXT: v_bfe_u32 v28, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v11
-; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28
-; VI-NEXT: v_or_b32_e32 v29, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v12
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_bfe_u32 v29, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v28
-; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v30, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc
-; VI-NEXT: v_bfe_u32 v29, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v12
-; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29
-; VI-NEXT: v_or_b32_e32 v30, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v13
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_bfe_u32 v30, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v29
-; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v31, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc
-; VI-NEXT: v_bfe_u32 v30, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v13
-; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30
-; VI-NEXT: v_or_b32_e32 v31, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v14
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_bfe_u32 v31, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v30
-; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v32, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc
-; VI-NEXT: v_bfe_u32 v31, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v14
-; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31
-; VI-NEXT: v_or_b32_e32 v32, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v15
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_bfe_u32 v32, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v31
-; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v33, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
-; VI-NEXT: v_bfe_u32 v32, v15, 16, 1
-; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v15
-; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32
-; VI-NEXT: v_or_b32_e32 v33, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v15, v15, v31, 16
-; VI-NEXT: v_alignbit_b32 v14, v14, v30, 16
-; VI-NEXT: v_alignbit_b32 v13, v13, v29, 16
-; VI-NEXT: v_alignbit_b32 v12, v12, v28, 16
-; VI-NEXT: v_alignbit_b32 v11, v11, v27, 16
-; VI-NEXT: v_alignbit_b32 v10, v10, v26, 16
-; VI-NEXT: v_alignbit_b32 v9, v9, v25, 16
-; VI-NEXT: v_alignbit_b32 v8, v8, v24, 16
-; VI-NEXT: v_alignbit_b32 v7, v7, v23, 16
-; VI-NEXT: v_alignbit_b32 v6, v6, v22, 16
-; VI-NEXT: v_alignbit_b32 v5, v5, v21, 16
-; VI-NEXT: v_alignbit_b32 v4, v4, v20, 16
-; VI-NEXT: v_alignbit_b32 v3, v3, v19, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v18, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v17, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB47_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB47_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_bfe_u32 v18, v17, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v18, v18, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc
-; GFX9-NEXT: v_bfe_u32 v18, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v18, v18, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v19, v19, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc
-; GFX9-NEXT: v_bfe_u32 v19, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v19, v19, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v20, v20, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc
-; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_bfe_u32 v21, v20, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v21, v21, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc
-; GFX9-NEXT: v_bfe_u32 v21, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v21, v21, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_bfe_u32 v22, v21, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v22, v22, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc
-; GFX9-NEXT: v_bfe_u32 v22, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v22, v22, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_bfe_u32 v23, v22, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v23, v23, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc
-; GFX9-NEXT: v_bfe_u32 v23, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v23, v23, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_bfe_u32 v24, v23, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v24, v24, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc
-; GFX9-NEXT: v_bfe_u32 v24, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v24, v24, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_bfe_u32 v25, v24, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v25, v25, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc
-; GFX9-NEXT: v_bfe_u32 v25, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v25, v25, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v9
-; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX9-NEXT: v_bfe_u32 v26, v25, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX9-NEXT: v_add3_u32 v26, v26, v25, s6
-; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v25
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc
-; GFX9-NEXT: v_bfe_u32 v26, v9, 16, 1
-; GFX9-NEXT: v_add3_u32 v26, v26, v9, s6
-; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v9
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; GFX9-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v10
-; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GFX9-NEXT: v_bfe_u32 v27, v26, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX9-NEXT: v_add3_u32 v27, v27, v26, s6
-; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v26
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc
-; GFX9-NEXT: v_bfe_u32 v27, v10, 16, 1
-; GFX9-NEXT: v_add3_u32 v27, v27, v10, s6
-; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v10
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; GFX9-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v11
-; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX9-NEXT: v_bfe_u32 v28, v27, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; GFX9-NEXT: v_add3_u32 v28, v28, v27, s6
-; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v27
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc
-; GFX9-NEXT: v_bfe_u32 v28, v11, 16, 1
-; GFX9-NEXT: v_add3_u32 v28, v28, v11, s6
-; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v11
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; GFX9-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v12
-; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX9-NEXT: v_bfe_u32 v29, v28, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX9-NEXT: v_add3_u32 v29, v29, v28, s6
-; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v28
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc
-; GFX9-NEXT: v_bfe_u32 v29, v12, 16, 1
-; GFX9-NEXT: v_add3_u32 v29, v29, v12, s6
-; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v12
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; GFX9-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v13
-; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GFX9-NEXT: v_bfe_u32 v30, v29, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX9-NEXT: v_add3_u32 v30, v30, v29, s6
-; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v29
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc
-; GFX9-NEXT: v_bfe_u32 v30, v13, 16, 1
-; GFX9-NEXT: v_add3_u32 v30, v30, v13, s6
-; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v13
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; GFX9-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v14
-; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GFX9-NEXT: v_bfe_u32 v31, v30, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX9-NEXT: v_add3_u32 v31, v31, v30, s6
-; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v30
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc
-; GFX9-NEXT: v_bfe_u32 v31, v14, 16, 1
-; GFX9-NEXT: v_add3_u32 v31, v31, v14, s6
-; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; GFX9-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v15
-; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX9-NEXT: v_add3_u32 v32, v32, v31, s6
-; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
-; GFX9-NEXT: v_bfe_u32 v32, v15, 16, 1
-; GFX9-NEXT: v_add3_u32 v32, v32, v15, s6
-; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v15
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; GFX9-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc
-; GFX9-NEXT: s_mov_b32 s6, 0x7060302
-; GFX9-NEXT: v_perm_b32 v15, v15, v31, s6
-; GFX9-NEXT: v_perm_b32 v14, v14, v30, s6
-; GFX9-NEXT: v_perm_b32 v13, v13, v29, s6
-; GFX9-NEXT: v_perm_b32 v12, v12, v28, s6
-; GFX9-NEXT: v_perm_b32 v11, v11, v27, s6
-; GFX9-NEXT: v_perm_b32 v10, v10, v26, s6
-; GFX9-NEXT: v_perm_b32 v9, v9, v25, s6
-; GFX9-NEXT: v_perm_b32 v8, v8, v24, s6
-; GFX9-NEXT: v_perm_b32 v7, v7, v23, s6
-; GFX9-NEXT: v_perm_b32 v6, v6, v22, s6
-; GFX9-NEXT: v_perm_b32 v5, v5, v21, s6
-; GFX9-NEXT: v_perm_b32 v4, v4, v20, s6
-; GFX9-NEXT: v_perm_b32 v3, v3, v19, s6
-; GFX9-NEXT: v_perm_b32 v2, v2, v18, s6
-; GFX9-NEXT: v_perm_b32 v1, v1, v17, s6
-; GFX9-NEXT: v_perm_b32 v0, v0, v16, s6
-; GFX9-NEXT: .LBB47_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32bf16_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB47_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v11
-; GFX11-NEXT: v_dual_add_f32 v24, 0x40c00000, v24 :: v_dual_lshlrev_b32 v25, 16, v6
-; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v19, v16, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v16
-; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff
-; GFX11-NEXT: v_add3_u32 v19, v19, v16, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11-NEXT: v_dual_add_f32 v26, 0x40c00000, v26 :: v_dual_lshlrev_b32 v27, 16, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v16, v19, v22, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v22, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v10
-; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-NEXT: v_dual_add_f32 v28, 0x40c00000, v28 :: v_dual_add_f32 v29, 0x40c00000, v29
-; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v31, 16, v12
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v0, v20, v23 :: v_dual_lshlrev_b32 v23, 16, v4
-; GFX11-NEXT: v_bfe_u32 v20, v1, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX11-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_add_f32 v23, 0x40c00000, v23
-; GFX11-NEXT: v_perm_b32 v0, v0, v16, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v17, v21, v19, vcc_lo
-; GFX11-NEXT: v_add3_u32 v19, v20, v1, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v31, 0x40c00000, v31
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v1, v19, v20 :: v_dual_add_f32 v18, 0x40c00000, v18
-; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-NEXT: v_bfe_u32 v21, v18, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v31
-; GFX11-NEXT: v_add3_u32 v19, v21, v18, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1
-; GFX11-NEXT: v_perm_b32 v1, v1, v17, 0x7060302
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_dual_cndmask_b32 v18, v19, v20 :: v_dual_and_b32 v7, 0xffff0000, v7
-; GFX11-NEXT: v_add3_u32 v19, v21, v2, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT: v_bfe_u32 v21, v22, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; GFX11-NEXT: v_dual_cndmask_b32 v2, v19, v20 :: v_dual_and_b32 v11, 0xffff0000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v19, v21, v22, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v22
-; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-NEXT: v_bfe_u32 v22, v23, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302
-; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-NEXT: v_cndmask_b32_e32 v19, v19, v20, vcc_lo
-; GFX11-NEXT: v_add3_u32 v20, v21, v3, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT: v_or_b32_e32 v32, 0x400000, v11
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v20, v22, v23, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v23
-; GFX11-NEXT: v_bfe_u32 v22, v4, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-NEXT: v_bfe_u32 v23, v24, 16, 1
-; GFX11-NEXT: v_perm_b32 v3, v3, v19, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v20, v20, v21, vcc_lo
-; GFX11-NEXT: v_add3_u32 v21, v22, v4, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc_lo
-; GFX11-NEXT: v_add3_u32 v21, v23, v24, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v24
-; GFX11-NEXT: v_bfe_u32 v23, v5, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-NEXT: v_bfe_u32 v24, v25, 16, 1
-; GFX11-NEXT: v_perm_b32 v4, v4, v20, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v21, v21, v22, vcc_lo
-; GFX11-NEXT: v_add3_u32 v22, v23, v5, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc_lo
-; GFX11-NEXT: v_add3_u32 v22, v24, v25, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v25
-; GFX11-NEXT: v_bfe_u32 v24, v6, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-NEXT: v_bfe_u32 v25, v26, 16, 1
-; GFX11-NEXT: v_perm_b32 v5, v5, v21, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v22, v22, v23, vcc_lo
-; GFX11-NEXT: v_add3_u32 v23, v24, v6, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc_lo
-; GFX11-NEXT: v_add3_u32 v23, v25, v26, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v26
-; GFX11-NEXT: v_bfe_u32 v25, v7, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-NEXT: v_bfe_u32 v26, v27, 16, 1
-; GFX11-NEXT: v_perm_b32 v6, v6, v22, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v23, v23, v24, vcc_lo
-; GFX11-NEXT: v_add3_u32 v24, v25, v7, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc_lo
-; GFX11-NEXT: v_add3_u32 v24, v26, v27, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v27
-; GFX11-NEXT: v_bfe_u32 v26, v8, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-NEXT: v_bfe_u32 v27, v28, 16, 1
-; GFX11-NEXT: v_perm_b32 v7, v7, v23, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v24, v24, v25, vcc_lo
-; GFX11-NEXT: v_add3_u32 v25, v26, v8, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v8
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc_lo
-; GFX11-NEXT: v_add3_u32 v25, v27, v28, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v28
-; GFX11-NEXT: v_bfe_u32 v27, v9, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: v_bfe_u32 v28, v29, 16, 1
-; GFX11-NEXT: v_perm_b32 v8, v8, v24, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v25, v25, v26, vcc_lo
-; GFX11-NEXT: v_add3_u32 v26, v27, v9, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v9
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc_lo
-; GFX11-NEXT: v_add3_u32 v26, v28, v29, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v29
-; GFX11-NEXT: v_bfe_u32 v28, v10, 16, 1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-NEXT: v_bfe_u32 v29, v30, 16, 1
-; GFX11-NEXT: v_perm_b32 v9, v9, v25, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v26, v26, v27, vcc_lo
-; GFX11-NEXT: v_add3_u32 v27, v28, v10, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v10
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc_lo
-; GFX11-NEXT: v_add3_u32 v27, v29, v30, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v30
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-NEXT: v_bfe_u32 v30, v31, 16, 1
-; GFX11-NEXT: v_bfe_u32 v29, v11, 16, 1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_dual_cndmask_b32 v27, v27, v28 :: v_dual_lshlrev_b32 v28, 16, v13
-; GFX11-NEXT: v_add3_u32 v30, v30, v31, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; GFX11-NEXT: v_add3_u32 v31, v34, v12, 0x7fff
-; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GFX11-NEXT: v_add3_u32 v29, v29, v11, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v30, v30, v33, vcc_lo
-; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v12
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-NEXT: v_bfe_u32 v35, v28, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v28
-; GFX11-NEXT: v_cndmask_b32_e32 v12, v31, v33, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v34, v35, v28, 0x7fff
-; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; GFX11-NEXT: v_bfe_u32 v37, v13, 16, 1
-; GFX11-NEXT: v_perm_b32 v10, v10, v26, 0x7060302
-; GFX11-NEXT: v_dual_add_f32 v31, 0x40c00000, v35 :: v_dual_cndmask_b32 v28, v34, v36
-; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v15
-; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-NEXT: v_add3_u32 v33, v37, v13, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT: v_bfe_u32 v35, v31, 16, 1
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34
-; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-NEXT: v_bfe_u32 v38, v14, 16, 1
-; GFX11-NEXT: v_add3_u32 v35, v35, v31, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-NEXT: v_bfe_u32 v39, v34, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v34
-; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v13
-; GFX11-NEXT: v_cndmask_b32_e32 v31, v35, v37, vcc_lo
-; GFX11-NEXT: v_add3_u32 v37, v38, v14, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-NEXT: v_bfe_u32 v35, v15, 16, 1
-; GFX11-NEXT: v_add3_u32 v39, v39, v34, 0x7fff
-; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v14, v37, v38, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-NEXT: v_add3_u32 v35, v35, v15, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_perm_b32 v14, v14, v31, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v34, v39, v48, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-NEXT: v_cndmask_b32_e32 v15, v35, v49, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_perm_b32 v15, v15, v34, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v36, vcc_lo
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-NEXT: v_perm_b32 v12, v12, v30, 0x7060302
-; GFX11-NEXT: v_perm_b32 v13, v13, v28, 0x7060302
-; GFX11-NEXT: v_cndmask_b32_e32 v11, v29, v32, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v11, v11, v27, 0x7060302
-; GFX11-NEXT: .LBB47_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0)
- %a2 = bitcast <32 x bfloat> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x bfloat> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32i16_to_v64i8:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v4
-; GCN-NEXT: s_waitcnt expcnt(6)
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v6
-; GCN-NEXT: s_waitcnt expcnt(4)
-; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v8
-; GCN-NEXT: s_waitcnt expcnt(3)
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v10
-; GCN-NEXT: s_waitcnt expcnt(2)
-; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v12
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v14
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v49
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; kill: killed $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v21
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v38
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v4
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v8
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v12
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v16
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v49
-; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: v_bfe_u32 v4, v4, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v8, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: v_bfe_u32 v44, v12, 8, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v16, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v20, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v24, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v28, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_bfe_u32 v4, v49, 8, 8
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: v_or_b32_e32 v52, v1, v46
-; GCN-NEXT: v_or_b32_e32 v51, v2, v47
-; GCN-NEXT: v_or_b32_e32 v48, v3, v57
-; GCN-NEXT: v_or_b32_e32 v37, v5, v59
-; GCN-NEXT: v_or_b32_e32 v32, v6, v60
-; GCN-NEXT: v_or_b32_e32 v6, v7, v61
-; GCN-NEXT: v_or_b32_e32 v58, v9, v62
-; GCN-NEXT: v_or_b32_e32 v56, v11, v63
-; GCN-NEXT: v_or_b32_e32 v42, v13, v10
-; GCN-NEXT: v_or_b32_e32 v54, v15, v14
-; GCN-NEXT: v_or_b32_e32 v50, v17, v22
-; GCN-NEXT: v_or_b32_e32 v39, v18, v33
-; GCN-NEXT: v_or_b32_e32 v36, v19, v26
-; GCN-NEXT: v_or_b32_e32 v31, v21, v34
-; GCN-NEXT: v_or_b32_e32 v18, v23, v30
-; GCN-NEXT: v_or_b32_e32 v2, v25, v35
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v55, v6, v32, 24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v32, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v6, v32, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v41, v56, v58, 24
-; GCN-NEXT: v_alignbit_b32 v43, v56, v58, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v56, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v45, v39, v50, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: .LBB48_2: ; %Flow
-; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v49, v40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_xor_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB48_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v29
-; GCN-NEXT: s_mov_b32 s6, 0x30000
-; GCN-NEXT: s_waitcnt vmcnt(3)
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v38
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v25
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v27
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v21
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v23
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v19
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_or_b32_e32 v2, v30, v2
-; GCN-NEXT: v_or_b32_e32 v4, v35, v4
-; GCN-NEXT: v_or_b32_e32 v6, v26, v6
-; GCN-NEXT: v_or_b32_e32 v8, v34, v8
-; GCN-NEXT: v_or_b32_e32 v12, v22, v12
-; GCN-NEXT: v_or_b32_e32 v16, v33, v16
-; GCN-NEXT: v_or_b32_e32 v10, v10, v17
-; GCN-NEXT: v_or_b32_e32 v14, v14, v18
-; GCN-NEXT: v_or_b32_e32 v13, v62, v13
-; GCN-NEXT: v_or_b32_e32 v15, v63, v15
-; GCN-NEXT: v_or_b32_e32 v9, v60, v9
-; GCN-NEXT: v_or_b32_e32 v11, v61, v11
-; GCN-NEXT: v_or_b32_e32 v5, v57, v5
-; GCN-NEXT: v_or_b32_e32 v7, v59, v7
-; GCN-NEXT: v_or_b32_e32 v1, v46, v1
-; GCN-NEXT: v_or_b32_e32 v3, v47, v3
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v2
-; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v31, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v16
-; GCN-NEXT: v_add_i32_e32 v42, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v58, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v56, vcc, s6, v15
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v3
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v4, v37, v48, 24
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v8, v37, v48, 8
-; GCN-NEXT: v_alignbit_b32 v55, v6, v32, 24
-; GCN-NEXT: v_alignbit_b32 v20, v6, v32, 16
-; GCN-NEXT: v_alignbit_b32 v12, v6, v32, 8
-; GCN-NEXT: v_alignbit_b32 v41, v56, v58, 24
-; GCN-NEXT: v_alignbit_b32 v43, v56, v58, 16
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v56, v58, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GCN-NEXT: v_alignbit_b32 v45, v39, v50, 8
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 24
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v51
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 24, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v37
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v6
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v56
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v54
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v39
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v31
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v2
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: .LBB48_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v52
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v5, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v51
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v7, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v8
-; GCN-NEXT: v_or_b32_e32 v8, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v10, v1, v3
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v12
-; GCN-NEXT: v_or_b32_e32 v12, v1, v3
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v1
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v28
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v1
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GCN-NEXT: v_or_b32_e32 v13, v1, v3
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v3
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v4
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v58
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GCN-NEXT: v_or_b32_e32 v14, v3, v4
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 8, v0
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v24
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v56
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v6
-; GCN-NEXT: v_or_b32_e32 v16, v4, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, 12, v0
-; GCN-NEXT: v_and_b32_e32 v38, 0xff, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v55
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v42
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GCN-NEXT: v_or_b32_e32 v17, v6, v9
-; GCN-NEXT: v_add_i32_e32 v6, vcc, 16, v0
-; GCN-NEXT: v_and_b32_e32 v48, 0xff, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v44
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v54
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GCN-NEXT: v_or_b32_e32 v21, v9, v11
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 20, v0
-; GCN-NEXT: v_and_b32_e32 v51, 0xff, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v41
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v45
-; GCN-NEXT: v_or_b32_e32 v24, v11, v15
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 24, v0
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v55, 0xff, v15
-; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v15
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v39
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GCN-NEXT: v_or_b32_e32 v22, v15, v19
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 28, v0
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v42, 0xff, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v19
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v36
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25
-; GCN-NEXT: v_or_b32_e32 v36, v19, v25
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 32, v0
-; GCN-NEXT: v_and_b32_e32 v45, 0xff, v49
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v25
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v31
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31
-; GCN-NEXT: v_or_b32_e32 v31, v25, v31
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 36, v0
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v46, 0xff, v50
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v52, 24, v50
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v50
-; GCN-NEXT: v_or_b32_e32 v50, v18, v50
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 40, v0
-; GCN-NEXT: v_and_b32_e32 v47, 0xff, v53
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v53, 24, v53
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v54
-; GCN-NEXT: v_or_b32_e32 v54, v2, v54
-; GCN-NEXT: v_add_i32_e32 v2, vcc, 44, v0
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v56, 0xff, v40
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GCN-NEXT: v_or_b32_e32 v33, v33, v32
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 48, v0
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v57, 0xff, v41
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; GCN-NEXT: v_or_b32_e32 v43, v27, v26
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 52, v0
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v58, 0xff, v27
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 24, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; GCN-NEXT: v_or_b32_e32 v29, v29, v27
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 56, v0
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v59, 0xff, v28
-; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v34
-; GCN-NEXT: v_or_b32_e32 v34, v35, v28
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 60, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v59
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_or_b32_e32 v20, v20, v35
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_or_b32_e32 v23, v23, v38
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_or_b32_e32 v30, v30, v48
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v35, v37, v51
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GCN-NEXT: v_or_b32_e32 v37, v39, v55
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GCN-NEXT: v_or_b32_e32 v38, v49, v42
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GCN-NEXT: v_or_b32_e32 v39, v52, v45
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GCN-NEXT: v_or_b32_e32 v48, v53, v46
-; GCN-NEXT: v_and_b32_e32 v36, 0xffff, v36
-; GCN-NEXT: v_or_b32_e32 v49, v40, v47
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GCN-NEXT: v_or_b32_e32 v51, v41, v56
-; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GCN-NEXT: v_or_b32_e32 v52, v44, v57
-; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v54
-; GCN-NEXT: v_or_b32_e32 v54, v60, v58
-; GCN-NEXT: v_or_b32_e32 v5, v5, v33
-; GCN-NEXT: v_or_b32_e32 v7, v7, v43
-; GCN-NEXT: v_or_b32_e32 v8, v8, v29
-; GCN-NEXT: v_or_b32_e32 v10, v10, v34
-; GCN-NEXT: v_or_b32_e32 v12, v12, v20
-; GCN-NEXT: v_or_b32_e32 v13, v13, v23
-; GCN-NEXT: v_or_b32_e32 v14, v14, v30
-; GCN-NEXT: v_or_b32_e32 v16, v16, v35
-; GCN-NEXT: v_or_b32_e32 v17, v17, v37
-; GCN-NEXT: v_or_b32_e32 v20, v21, v38
-; GCN-NEXT: v_or_b32_e32 v21, v24, v39
-; GCN-NEXT: v_or_b32_e32 v22, v22, v48
-; GCN-NEXT: v_or_b32_e32 v23, v36, v49
-; GCN-NEXT: v_or_b32_e32 v24, v31, v51
-; GCN-NEXT: v_or_b32_e32 v29, v50, v52
-; GCN-NEXT: v_or_b32_e32 v30, v53, v54
-; GCN-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v7, v1, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v8, v3, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v10, v4, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v12, v6, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v13, v9, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v14, v11, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v16, v15, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v17, v19, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v20, v25, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v21, v18, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v23, v32, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v24, v26, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v29, v27, s[0:3], 0 offen
-; GCN-NEXT: buffer_store_dword v30, v28, s[0:3], 0 offen
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32i16_to_v64i8:
-; VI: ; %bb.0:
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v16
-; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
-; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v1
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: ; kill: killed $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; kill: killed $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v15
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v14
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v14
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v13
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v12
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v12
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v11
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v10
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v8
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v8
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v7
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v6
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[15:16]
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v5
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[13:14]
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v16
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6
-; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v4
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v2
-; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v2
-; VI-NEXT: v_lshrrev_b32_e32 v18, 8, v1
-; VI-NEXT: v_mov_b32_e32 v25, v50
-; VI-NEXT: v_mov_b32_e32 v41, v1
-; VI-NEXT: v_mov_b32_e32 v54, v2
-; VI-NEXT: v_mov_b32_e32 v57, v3
-; VI-NEXT: v_mov_b32_e32 v47, v4
-; VI-NEXT: v_mov_b32_e32 v61, v5
-; VI-NEXT: v_mov_b32_e32 v60, v6
-; VI-NEXT: v_mov_b32_e32 v52, v7
-; VI-NEXT: v_mov_b32_e32 v63, v8
-; VI-NEXT: v_mov_b32_e32 v40, v9
-; VI-NEXT: v_mov_b32_e32 v53, v10
-; VI-NEXT: v_mov_b32_e32 v17, v11
-; VI-NEXT: v_mov_b32_e32 v44, v12
-; VI-NEXT: v_mov_b32_e32 v58, v13
-; VI-NEXT: v_mov_b32_e32 v56, v14
-; VI-NEXT: v_mov_b32_e32 v50, v15
-; VI-NEXT: v_mov_b32_e32 v62, v16
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: ; implicit-def: $vgpr3
-; VI-NEXT: ; implicit-def: $vgpr5
-; VI-NEXT: ; implicit-def: $vgpr7
-; VI-NEXT: ; implicit-def: $vgpr9
-; VI-NEXT: ; implicit-def: $vgpr11
-; VI-NEXT: ; implicit-def: $vgpr13
-; VI-NEXT: ; implicit-def: $vgpr15
-; VI-NEXT: .LBB48_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB48_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v18, 3
-; VI-NEXT: v_add_u16_sdwa v26, v16, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v29, v15, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v62, 3, v16
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v26
-; VI-NEXT: v_add_u16_e32 v50, 3, v15
-; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v29
-; VI-NEXT: v_or_b32_e32 v16, v62, v16
-; VI-NEXT: v_or_b32_e32 v15, v50, v15
-; VI-NEXT: v_add_u16_sdwa v38, v2, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v49, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v36, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v48, v3, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v34, v6, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v39, v5, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v32, v8, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v37, v7, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v30, v10, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v35, v9, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v28, v12, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v33, v11, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v27, v14, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_u16_sdwa v31, v13, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[15:16]
-; VI-NEXT: v_add_u16_e32 v56, 3, v14
-; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v27
-; VI-NEXT: v_add_u16_e32 v58, 3, v13
-; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v31
-; VI-NEXT: v_or_b32_e32 v14, v56, v14
-; VI-NEXT: v_or_b32_e32 v13, v58, v13
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; VI-NEXT: v_add_u16_e32 v44, 3, v12
-; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v28
-; VI-NEXT: v_add_u16_e32 v17, 3, v11
-; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v33
-; VI-NEXT: v_add_u16_e32 v53, 3, v10
-; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v30
-; VI-NEXT: v_add_u16_e32 v40, 3, v9
-; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v35
-; VI-NEXT: v_or_b32_e32 v12, v44, v12
-; VI-NEXT: v_or_b32_e32 v11, v17, v11
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; VI-NEXT: v_add_u16_e32 v63, 3, v8
-; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v32
-; VI-NEXT: v_add_u16_e32 v52, 3, v7
-; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v37
-; VI-NEXT: v_or_b32_e32 v10, v53, v10
-; VI-NEXT: v_or_b32_e32 v9, v40, v9
-; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; VI-NEXT: v_add_u16_e32 v60, 3, v6
-; VI-NEXT: v_lshlrev_b32_e32 v6, 16, v34
-; VI-NEXT: v_add_u16_e32 v61, 3, v5
-; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v39
-; VI-NEXT: v_or_b32_e32 v8, v63, v8
-; VI-NEXT: v_or_b32_e32 v7, v52, v7
-; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; VI-NEXT: v_add_u16_e32 v47, 3, v4
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v36
-; VI-NEXT: v_add_u16_e32 v57, 3, v3
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v48
-; VI-NEXT: v_or_b32_e32 v6, v60, v6
-; VI-NEXT: v_or_b32_e32 v5, v61, v5
-; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; VI-NEXT: v_add_u16_e32 v54, 3, v2
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v38
-; VI-NEXT: v_add_u16_e32 v41, 3, v1
-; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v49
-; VI-NEXT: v_or_b32_e32 v4, v47, v4
-; VI-NEXT: v_or_b32_e32 v3, v57, v3
-; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; VI-NEXT: v_or_b32_e32 v2, v54, v2
-; VI-NEXT: v_or_b32_e32 v1, v41, v1
-; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; VI-NEXT: v_lshrrev_b32_e32 v18, 8, v1
-; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; VI-NEXT: v_bfe_u32 v1, v27, 8, 8
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v13
-; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v9
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v28, 8, 8
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v12
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v7
-; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v6
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v30, 8, 8
-; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v16
-; VI-NEXT: v_lshrrev_b32_e32 v15, 8, v15
-; VI-NEXT: v_lshrrev_b32_e32 v14, 8, v14
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v11
-; VI-NEXT: v_lshrrev_b32_e32 v11, 8, v10
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v5
-; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v4
-; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v3
-; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v2
-; VI-NEXT: v_bfe_u32 v25, v26, 8, 8
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; VI-NEXT: v_bfe_u32 v1, v32, 8, 8
-; VI-NEXT: v_bfe_u32 v43, v34, 8, 8
-; VI-NEXT: v_bfe_u32 v46, v36, 8, 8
-; VI-NEXT: v_bfe_u32 v59, v38, 8, 8
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; VI-NEXT: .LBB48_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v18
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24
-; VI-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v45
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59
-; VI-NEXT: v_or_b32_sdwa v1, v54, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v51
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v23
-; VI-NEXT: v_or_b32_sdwa v1, v57, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v55
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v46
-; VI-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22
-; VI-NEXT: v_or_b32_sdwa v2, v39, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43
-; VI-NEXT: v_or_b32_sdwa v2, v34, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; VI-NEXT: v_or_b32_sdwa v2, v37, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v32, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v42
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; VI-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v53, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v44, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v31, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v27, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; VI-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v29, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0
-; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; VI-NEXT: v_or_b32_sdwa v2, v26, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; VI-NEXT: v_or_b32_sdwa v1, v62, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32i16_to_v64i8:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; kill: killed $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_mov_b32_e32 v26, v23
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: .LBB48_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB48_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4]
-; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2]
-; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16
-; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15
-; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14
-; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13
-; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12
-; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3
-; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: .LBB48_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25
-; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21
-; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50
-; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32
-; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32i16_to_v64i8:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB48_2: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB48_4
-; GFX11-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15
-; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14
-; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13
-; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12
-; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11
-; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10
-; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9
-; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8
-; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7
-; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6
-; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5
-; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2
-; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1
-; GFX11-NEXT: .LBB48_4: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25
-; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v25
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87
-; GFX11-NEXT: v_or_b32_e32 v24, v96, v24
-; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v25
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-NEXT: v_or_b32_e32 v25, v86, v85
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84
-; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v24
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v24
-; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81
-; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v84
-; GFX11-NEXT: v_or_b32_e32 v23, v83, v23
-; GFX11-NEXT: v_or_b32_e32 v25, v25, v80
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v71
-; GFX11-NEXT: v_or_b32_e32 v22, v70, v22
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v24
-; GFX11-NEXT: v_or_b32_e32 v3, v3, v23
-; GFX11-NEXT: v_or_b32_e32 v4, v4, v25
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-NEXT: v_or_b32_e32 v5, v5, v22
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-NEXT: v_or_b32_e32 v21, v23, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v24
-; GFX11-NEXT: v_or_b32_e32 v22, v25, v54
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52
-; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48
-; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-NEXT: v_or_b32_e32 v20, v23, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v24
-; GFX11-NEXT: v_or_b32_e32 v23, v25, v49
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v48
-; GFX11-NEXT: v_or_b32_e32 v19, v39, v19
-; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v21
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v22
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v20
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v23
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v19
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34
-; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69
-; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68
-; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v19, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v22
-; GFX11-NEXT: v_or_b32_e32 v18, v23, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v24
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28
-; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27
-; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v69
-; GFX11-NEXT: v_or_b32_e32 v67, v68, v67
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v22
-; GFX11-NEXT: v_or_b32_e32 v17, v23, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v24
-; GFX11-NEXT: v_or_b32_e32 v21, v25, v26
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v67
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v19
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v18
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v20
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v17
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v21
-; GFX11-NEXT: s_clause 0x3
-; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <32 x i16> %a, splat (i16 3)
- %a2 = bitcast <32 x i16> %a1 to <64 x i8>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x i16> %a to <64 x i8>
- br label %end
-
-end:
- %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <64 x i8> %phi
-}
-
-define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
-; GCN-LABEL: bitcast_v64i8_to_v32i16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(1)
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:128
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:92
-; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84
-; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68
-; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:64
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:60
-; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:56
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:52
-; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:48
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:44
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
-; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:132
-; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:120
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116
-; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:112
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:108
-; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:104
-; GCN-NEXT: s_waitcnt vmcnt(5)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39
-; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:28
-; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:24
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:20
-; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:16
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:12
-; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:8
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v13
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v15
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v11
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v21
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v23
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v19
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v29
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v27
-; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v9
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v17
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v38
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt vmcnt(14) expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v51
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v39
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v52
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v36
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v35
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v37
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v33
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v16
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v34
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v57, 8, v48
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v49
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v53
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:36
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:4
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:100
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:124
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v32
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v50
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v1, v1, v3
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v12
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v3, v3, v4
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v20
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v4, v5
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v28
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v5, v5, v7
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v56
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v46
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v47
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v62
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v61
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v41
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v63
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v40
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v58
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v33, 0xff, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v34, 0xff, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v35, 0xff, v0
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v0
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v59
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v60
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v42
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v8
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v6
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v7
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20
-; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v21, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v23
-; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v25, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v27
-; GCN-NEXT: v_or_b32_e32 v23, v28, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30
-; GCN-NEXT: v_or_b32_e32 v24, v24, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v33, v29
-; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v34, v30
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v35, v31
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v12, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v32
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v14, v32
-; GCN-NEXT: v_or_b32_e32 v15, v15, v44
-; GCN-NEXT: v_or_b32_e32 v16, v16, v45
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v40, v32, v11
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v11, v8
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v11, v9
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v11, v6
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v41, v11, v7
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v7, v0
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v7, v2
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v17
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v11, v18
-; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v11, v19
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v20
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v19, v21
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v19, v22
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v23
-; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v22, v25
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v23, v26
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24
-; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v24, v27
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v28
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v29
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v30
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v31
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_or_b32_e32 v33, v1, v10
-; GCN-NEXT: v_or_b32_e32 v35, v3, v8
-; GCN-NEXT: v_or_b32_e32 v37, v4, v6
-; GCN-NEXT: v_or_b32_e32 v39, v5, v0
-; GCN-NEXT: v_or_b32_e32 v49, v7, v17
-; GCN-NEXT: v_or_b32_e32 v51, v11, v20
-; GCN-NEXT: v_or_b32_e32 v53, v19, v22
-; GCN-NEXT: v_or_b32_e32 v55, v23, v24
-; GCN-NEXT: v_or_b32_e32 v32, v27, v40
-; GCN-NEXT: v_or_b32_e32 v34, v28, v9
-; GCN-NEXT: v_or_b32_e32 v36, v29, v41
-; GCN-NEXT: v_or_b32_e32 v38, v12, v2
-; GCN-NEXT: v_or_b32_e32 v48, v13, v18
-; GCN-NEXT: v_or_b32_e32 v50, v14, v21
-; GCN-NEXT: v_or_b32_e32 v52, v15, v25
-; GCN-NEXT: v_or_b32_e32 v54, v16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v22
-; GCN-NEXT: v_alignbit_b32 v1, v33, v40, 16
-; GCN-NEXT: v_alignbit_b32 v5, v35, v9, 16
-; GCN-NEXT: v_alignbit_b32 v9, v37, v41, 16
-; GCN-NEXT: v_alignbit_b32 v13, v39, v2, 16
-; GCN-NEXT: v_alignbit_b32 v17, v49, v18, 16
-; GCN-NEXT: v_alignbit_b32 v21, v51, v21, 16
-; GCN-NEXT: v_alignbit_b32 v25, v53, v25, 16
-; GCN-NEXT: v_alignbit_b32 v29, v55, v26, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; kill: killed $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; kill: killed $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; kill: killed $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: .LBB49_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB49_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: s_waitcnt vmcnt(2) expcnt(0)
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v8
-; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GCN-NEXT: v_or_b32_e32 v1, v45, v1
-; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v24
-; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GCN-NEXT: v_or_b32_e32 v3, v57, v3
-; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v42
-; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GCN-NEXT: v_or_b32_e32 v5, v44, v5
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v41
-; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GCN-NEXT: v_or_b32_e32 v7, v43, v7
-; GCN-NEXT: s_movk_i32 s7, 0x300
-; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v58
-; GCN-NEXT: s_mov_b32 s6, 0x3000000
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v16
-; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v40
-; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v63
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v60
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v61
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v62
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v59
-; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v47
-; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56
-; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v46
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v26
-; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v28
-; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v30
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v0
-; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v2
-; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v4
-; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v6
-; GCN-NEXT: v_and_b32_e32 v0, 0xff, v32
-; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GCN-NEXT: v_and_b32_e32 v4, 0xff, v34
-; GCN-NEXT: v_and_b32_e32 v6, 0xff, v35
-; GCN-NEXT: v_and_b32_e32 v32, 0xff, v8
-; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19
-; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30
-; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v2
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v4, v4, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v9
-; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v9, v9, v11
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v13
-; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v13, v13, v15
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v16
-; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v16, v16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v19
-; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v19, v19, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v23
-; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v23, v23, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v25
-; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v25, v25, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v20, v26, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v26, v26, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v12, v27, v12
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v27, v27, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v29
-; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v29, v29, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v31
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v31, v31, v33
-; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v8, v32, v8
-; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v2, v32, v2
-; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v0, v32, v0
-; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v6, v32, v6
-; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v11, v32, v11
-; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v15, v32, v15
-; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v17, v32, v17
-; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v21, v32, v21
-; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v24, v32, v24
-; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v18, v32, v18
-; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v22, v32, v22
-; GCN-NEXT: v_add_i32_e32 v26, vcc, s7, v26
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v10, v32, v10
-; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v14, v32, v14
-; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v28, v32, v28
-; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29
-; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_or_b32_e32 v30, v32, v30
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GCN-NEXT: v_or_b32_e32 v1, v31, v1
-; GCN-NEXT: v_or_b32_e32 v3, v8, v3
-; GCN-NEXT: v_or_b32_e32 v2, v2, v5
-; GCN-NEXT: v_or_b32_e32 v0, v0, v7
-; GCN-NEXT: v_or_b32_e32 v4, v6, v4
-; GCN-NEXT: v_or_b32_e32 v5, v11, v9
-; GCN-NEXT: v_or_b32_e32 v6, v15, v13
-; GCN-NEXT: v_or_b32_e32 v7, v17, v16
-; GCN-NEXT: v_or_b32_e32 v8, v21, v19
-; GCN-NEXT: v_or_b32_e32 v9, v24, v23
-; GCN-NEXT: v_or_b32_e32 v11, v18, v25
-; GCN-NEXT: v_or_b32_e32 v13, v22, v20
-; GCN-NEXT: v_or_b32_e32 v10, v10, v26
-; GCN-NEXT: v_or_b32_e32 v12, v14, v12
-; GCN-NEXT: v_or_b32_e32 v14, v28, v27
-; GCN-NEXT: v_or_b32_e32 v15, v30, v29
-; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v1
-; GCN-NEXT: v_add_i32_e32 v55, vcc, s6, v3
-; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v2
-; GCN-NEXT: v_add_i32_e32 v53, vcc, s6, v0
-; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v4
-; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v5
-; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v6
-; GCN-NEXT: v_add_i32_e32 v49, vcc, s6, v7
-; GCN-NEXT: v_add_i32_e32 v38, vcc, s6, v8
-; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v9
-; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v11
-; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v13
-; GCN-NEXT: v_add_i32_e32 v34, vcc, s6, v10
-; GCN-NEXT: v_add_i32_e32 v35, vcc, s6, v12
-; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v14
-; GCN-NEXT: v_add_i32_e32 v33, vcc, s6, v15
-; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 16
-; GCN-NEXT: v_alignbit_b32 v5, v35, v34, 16
-; GCN-NEXT: v_alignbit_b32 v9, v37, v36, 16
-; GCN-NEXT: v_alignbit_b32 v13, v39, v38, 16
-; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16
-; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16
-; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16
-; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55
-; GCN-NEXT: .LBB49_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: v_mov_b32_e32 v0, v32
-; GCN-NEXT: v_mov_b32_e32 v2, v33
-; GCN-NEXT: v_mov_b32_e32 v4, v34
-; GCN-NEXT: v_mov_b32_e32 v6, v35
-; GCN-NEXT: s_waitcnt vmcnt(2)
-; GCN-NEXT: v_mov_b32_e32 v8, v36
-; GCN-NEXT: v_mov_b32_e32 v10, v37
-; GCN-NEXT: v_mov_b32_e32 v12, v38
-; GCN-NEXT: v_mov_b32_e32 v14, v39
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_mov_b32_e32 v16, v48
-; GCN-NEXT: v_mov_b32_e32 v18, v49
-; GCN-NEXT: v_mov_b32_e32 v20, v50
-; GCN-NEXT: v_mov_b32_e32 v22, v51
-; GCN-NEXT: v_mov_b32_e32 v24, v52
-; GCN-NEXT: v_mov_b32_e32 v26, v53
-; GCN-NEXT: v_mov_b32_e32 v28, v54
-; GCN-NEXT: v_mov_b32_e32 v30, v55
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v64i8_to_v32i16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124
-; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v1
-; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v5
-; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v7
-; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v9
-; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v11
-; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v13
-; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v15
-; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17
-; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19
-; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23
-; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
-; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v0
-; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v2
-; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v4
-; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v8
-; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v10
-; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v12
-; VI-NEXT: v_lshlrev_b16_e32 v60, 8, v14
-; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v16
-; VI-NEXT: v_lshlrev_b16_e32 v63, 8, v18
-; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v20
-; VI-NEXT: s_waitcnt vmcnt(13)
-; VI-NEXT: v_lshlrev_b16_e32 v24, 8, v22
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(11)
-; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v28
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v30
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v31
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:108
-; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:84
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:52
-; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:28
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v32
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr39
-; VI-NEXT: ; implicit-def: $vgpr20
-; VI-NEXT: ; implicit-def: $vgpr28
-; VI-NEXT: ; implicit-def: $vgpr22
-; VI-NEXT: ; implicit-def: $vgpr30
-; VI-NEXT: ; implicit-def: $vgpr35
-; VI-NEXT: ; implicit-def: $vgpr31
-; VI-NEXT: ; implicit-def: $vgpr38
-; VI-NEXT: ; implicit-def: $vgpr47
-; VI-NEXT: ; implicit-def: $vgpr57
-; VI-NEXT: ; implicit-def: $vgpr59
-; VI-NEXT: ; implicit-def: $vgpr62
-; VI-NEXT: ; implicit-def: $vgpr18
-; VI-NEXT: ; implicit-def: $vgpr33
-; VI-NEXT: ; implicit-def: $vgpr32
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr50
-; VI-NEXT: ; implicit-def: $vgpr51
-; VI-NEXT: ; implicit-def: $vgpr52
-; VI-NEXT: ; implicit-def: $vgpr53
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr19
-; VI-NEXT: ; implicit-def: $vgpr23
-; VI-NEXT: ; implicit-def: $vgpr27
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr45
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr54
-; VI-NEXT: ; implicit-def: $vgpr40
-; VI-NEXT: ; implicit-def: $vgpr42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr41
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr17
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr21
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: ; implicit-def: $vgpr25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; kill: killed $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr49
-; VI-NEXT: ; implicit-def: $vgpr48
-; VI-NEXT: ; implicit-def: $vgpr55
-; VI-NEXT: ; implicit-def: $vgpr36
-; VI-NEXT: ; implicit-def: $vgpr43
-; VI-NEXT: ; implicit-def: $vgpr37
-; VI-NEXT: ; implicit-def: $vgpr44
-; VI-NEXT: ; implicit-def: $vgpr61
-; VI-NEXT: ; implicit-def: $vgpr29
-; VI-NEXT: ; implicit-def: $vgpr46
-; VI-NEXT: ; implicit-def: $vgpr56
-; VI-NEXT: ; implicit-def: $vgpr58
-; VI-NEXT: ; implicit-def: $vgpr60
-; VI-NEXT: ; implicit-def: $vgpr63
-; VI-NEXT: ; implicit-def: $vgpr16
-; VI-NEXT: ; implicit-def: $vgpr24
-; VI-NEXT: ; implicit-def: $vgpr26
-; VI-NEXT: ; implicit-def: $vgpr34
-; VI-NEXT: .LBB49_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB49_4
-; VI-NEXT: ; %bb.3: ; %cmp.true
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_add_u16_e32 v0, 3, v38
-; VI-NEXT: v_add_u16_e32 v2, 3, v44
-; VI-NEXT: v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u16_e32 v2, 3, v35
-; VI-NEXT: v_mov_b32_e32 v3, 0x300
-; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_add_u16_e32 v0, 3, v37
-; VI-NEXT: v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(4)
-; VI-NEXT: v_add_u16_e32 v0, 3, v30
-; VI-NEXT: v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v43
-; VI-NEXT: v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_add_u16_e32 v0, 3, v22
-; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u16_e32 v0, 3, v36
-; VI-NEXT: v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u16_e32 v0, 3, v28
-; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v55
-; VI-NEXT: v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v20
-; VI-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v48
-; VI-NEXT: v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v0, 3, v39
-; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v0, 3, v49
-; VI-NEXT: v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v1, 3, v61
-; VI-NEXT: v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v1, 3, v31
-; VI-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT: v_or_b32_e32 v12, v16, v12
-; VI-NEXT: v_add_u16_e32 v16, 0x300, v24
-; VI-NEXT: v_add_u16_e32 v14, 0x300, v14
-; VI-NEXT: v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT: v_or_b32_e32 v13, v16, v13
-; VI-NEXT: v_or_b32_e32 v14, v14, v26
-; VI-NEXT: v_or_b32_e32 v15, v15, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v25, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v5, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v21, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v17, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v17
-; VI-NEXT: v_or_b32_e32 v4, v17, v4
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v21
-; VI-NEXT: v_or_b32_e32 v5, v17, v5
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v25
-; VI-NEXT: v_or_b32_e32 v6, v17, v6
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v29
-; VI-NEXT: v_or_b32_e32 v7, v17, v7
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v30
-; VI-NEXT: v_or_b32_e32 v8, v17, v8
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v20
-; VI-NEXT: v_or_b32_e32 v9, v17, v9
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v28
-; VI-NEXT: v_or_b32_e32 v10, v17, v10
-; VI-NEXT: v_add_u16_e32 v17, 0x300, v22
-; VI-NEXT: v_or_b32_e32 v11, v17, v11
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v0, 3, v0
-; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u16_e32 v3, 3, v3
-; VI-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v3
-; VI-NEXT: v_or_b32_e32 v0, v3, v0
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v31
-; VI-NEXT: v_or_b32_e32 v1, v3, v1
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v27
-; VI-NEXT: v_or_b32_e32 v2, v3, v2
-; VI-NEXT: v_add_u16_e32 v3, 0x300, v23
-; VI-NEXT: v_or_b32_e32 v3, v3, v19
-; VI-NEXT: .LBB49_4: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v64i8_to_v32i16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:112
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:120
-; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:36
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:4
-; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v3
-; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v5
-; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v7
-; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v13
-; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v15
-; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v17
-; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v19
-; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25
-; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v0
-; GFX9-NEXT: s_waitcnt vmcnt(22)
-; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v2
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v4
-; GFX9-NEXT: s_waitcnt vmcnt(20)
-; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(19)
-; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(18)
-; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(17)
-; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b16_e32 v60, 8, v16
-; GFX9-NEXT: s_waitcnt vmcnt(14)
-; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v18
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v20
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_lshlrev_b16_e32 v30, 8, v24
-; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:116
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28
-; GFX9-NEXT: s_waitcnt vmcnt(10)
-; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v31
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v32
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v26
-; GFX9-NEXT: s_waitcnt vmcnt(16)
-; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(9)
-; GFX9-NEXT: v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr39
-; GFX9-NEXT: ; implicit-def: $vgpr62
-; GFX9-NEXT: ; implicit-def: $vgpr63
-; GFX9-NEXT: ; implicit-def: $vgpr35
-; GFX9-NEXT: ; implicit-def: $vgpr34
-; GFX9-NEXT: ; implicit-def: $vgpr32
-; GFX9-NEXT: ; implicit-def: $vgpr38
-; GFX9-NEXT: ; implicit-def: $vgpr45
-; GFX9-NEXT: ; implicit-def: $vgpr56
-; GFX9-NEXT: ; implicit-def: $vgpr58
-; GFX9-NEXT: ; implicit-def: $vgpr60
-; GFX9-NEXT: ; implicit-def: $vgpr36
-; GFX9-NEXT: ; implicit-def: $vgpr26
-; GFX9-NEXT: ; implicit-def: $vgpr20
-; GFX9-NEXT: ; implicit-def: $vgpr33
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6
-; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr49
-; GFX9-NEXT: ; implicit-def: $vgpr48
-; GFX9-NEXT: ; implicit-def: $vgpr53
-; GFX9-NEXT: ; implicit-def: $vgpr50
-; GFX9-NEXT: ; implicit-def: $vgpr40
-; GFX9-NEXT: ; implicit-def: $vgpr17
-; GFX9-NEXT: ; implicit-def: $vgpr19
-; GFX9-NEXT: ; implicit-def: $vgpr23
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: ; implicit-def: $vgpr27
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr55
-; GFX9-NEXT: ; implicit-def: $vgpr54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr41
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr43
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr21
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: ; implicit-def: $vgpr25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT: v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT: v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT: v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT: v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT: v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT: v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT: v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT: v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; kill: killed $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr51
-; GFX9-NEXT: ; implicit-def: $vgpr52
-; GFX9-NEXT: ; implicit-def: $vgpr42
-; GFX9-NEXT: ; implicit-def: $vgpr37
-; GFX9-NEXT: ; implicit-def: $vgpr44
-; GFX9-NEXT: ; implicit-def: $vgpr28
-; GFX9-NEXT: ; implicit-def: $vgpr16
-; GFX9-NEXT: ; implicit-def: $vgpr47
-; GFX9-NEXT: ; implicit-def: $vgpr18
-; GFX9-NEXT: ; implicit-def: $vgpr29
-; GFX9-NEXT: ; implicit-def: $vgpr46
-; GFX9-NEXT: ; implicit-def: $vgpr57
-; GFX9-NEXT: ; implicit-def: $vgpr59
-; GFX9-NEXT: ; implicit-def: $vgpr61
-; GFX9-NEXT: ; implicit-def: $vgpr22
-; GFX9-NEXT: ; implicit-def: $vgpr30
-; GFX9-NEXT: ; implicit-def: $vgpr24
-; GFX9-NEXT: ; implicit-def: $vgpr31
-; GFX9-NEXT: .LBB49_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB49_4
-; GFX9-NEXT: ; %bb.3: ; %cmp.true
-; GFX9-NEXT: s_waitcnt vmcnt(8)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(5)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT: v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT: v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v44
-; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v37
-; GFX9-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v35
-; GFX9-NEXT: v_add_u16_e32 v2, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v42
-; GFX9-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v63
-; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v52
-; GFX9-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v62
-; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v51
-; GFX9-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v39
-; GFX9-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u16_e32 v3, 3, v32
-; GFX9-NEXT: v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v3
-; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v2
-; GFX9-NEXT: v_add_u16_e32 v1, 3, v38
-; GFX9-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v1
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
-; GFX9-NEXT: v_perm_b32 v8, v30, v8, s6
-; GFX9-NEXT: v_perm_b32 v9, v28, v9, s6
-; GFX9-NEXT: v_perm_b32 v10, v26, v10, s6
-; GFX9-NEXT: v_perm_b32 v11, v24, v11, s6
-; GFX9-NEXT: v_perm_b32 v12, v22, v12, s6
-; GFX9-NEXT: v_perm_b32 v13, v16, v13, s6
-; GFX9-NEXT: v_perm_b32 v14, v20, v14, s6
-; GFX9-NEXT: v_perm_b32 v15, v18, v15, s6
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT: v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v4, v17, v4, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v3, v21, v3, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v2, v25, v2, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: v_perm_b32 v1, v29, v1, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0
-; GFX9-NEXT: v_perm_b32 v0, v31, v0, s6
-; GFX9-NEXT: .LBB49_4: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v64i8_to_v32i16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v32, v14 :: v_dual_mov_b32 v31, v8
-; GFX11-NEXT: v_dual_mov_b32 v34, v12 :: v_dual_mov_b32 v35, v10
-; GFX11-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-NEXT: v_dual_mov_b32 v37, v2 :: v_dual_mov_b32 v36, v0
-; GFX11-NEXT: s_clause 0x1f
-; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128
-; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:124
-; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120
-; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:116
-; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112
-; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:108
-; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104
-; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:100
-; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96
-; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:92
-; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88
-; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132
-; GFX11-NEXT: scratch_load_u16 v14, off, s32
-; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:8
-; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:16
-; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:24
-; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:32
-; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:40
-; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:48
-; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:56
-; GFX11-NEXT: scratch_load_u16 v115, off, s32 offset:64
-; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:72
-; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:80
-; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:84
-; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:76
-; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:68
-; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:60
-; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:52
-; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:44
-; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:36
-; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:28
-; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:20
-; GFX11-NEXT: s_clause 0x1
-; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:12
-; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:4
-; GFX11-NEXT: v_lshlrev_b16 v50, 8, v1
-; GFX11-NEXT: v_lshlrev_b16 v51, 8, v3
-; GFX11-NEXT: v_lshlrev_b16 v49, 8, v5
-; GFX11-NEXT: v_lshlrev_b16 v39, 8, v7
-; GFX11-NEXT: v_lshlrev_b16 v54, 8, v9
-; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11
-; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13
-; GFX11-NEXT: v_lshlrev_b16 v52, 8, v15
-; GFX11-NEXT: v_lshlrev_b16 v53, 8, v17
-; GFX11-NEXT: v_lshlrev_b16 v17, 8, v19
-; GFX11-NEXT: v_lshlrev_b16 v83, 8, v21
-; GFX11-NEXT: v_lshlrev_b16 v81, 8, v23
-; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25
-; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27
-; GFX11-NEXT: v_lshlrev_b16 v27, 8, v29
-; GFX11-NEXT: s_waitcnt vmcnt(33)
-; GFX11-NEXT: v_lshlrev_b16 v119, 8, v0
-; GFX11-NEXT: s_waitcnt vmcnt(31)
-; GFX11-NEXT: v_lshlrev_b16 v118, 8, v2
-; GFX11-NEXT: s_waitcnt vmcnt(22)
-; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: s_waitcnt vmcnt(21)
-; GFX11-NEXT: v_lshlrev_b16 v23, 8, v14
-; GFX11-NEXT: s_waitcnt vmcnt(20)
-; GFX11-NEXT: v_lshlrev_b16 v98, 8, v96
-; GFX11-NEXT: s_waitcnt vmcnt(19)
-; GFX11-NEXT: v_lshlrev_b16 v25, 8, v100
-; GFX11-NEXT: s_waitcnt vmcnt(18)
-; GFX11-NEXT: v_lshlrev_b16 v96, 8, v101
-; GFX11-NEXT: s_waitcnt vmcnt(17)
-; GFX11-NEXT: v_lshlrev_b16 v29, 8, v102
-; GFX11-NEXT: s_waitcnt vmcnt(16)
-; GFX11-NEXT: v_lshlrev_b16 v112, 8, v103
-; GFX11-NEXT: s_waitcnt vmcnt(15)
-; GFX11-NEXT: v_lshlrev_b16 v103, 8, v113
-; GFX11-NEXT: s_waitcnt vmcnt(14)
-; GFX11-NEXT: v_lshlrev_b16 v101, 8, v114
-; GFX11-NEXT: s_waitcnt vmcnt(13)
-; GFX11-NEXT: v_lshlrev_b16 v100, 8, v115
-; GFX11-NEXT: s_waitcnt vmcnt(12)
-; GFX11-NEXT: v_lshlrev_b16 v114, 8, v116
-; GFX11-NEXT: s_waitcnt vmcnt(11)
-; GFX11-NEXT: v_lshlrev_b16 v102, 8, v117
-; GFX11-NEXT: v_lshlrev_b16 v117, 8, v10
-; GFX11-NEXT: v_lshlrev_b16 v113, 8, v8
-; GFX11-NEXT: v_lshlrev_b16 v116, 8, v6
-; GFX11-NEXT: v_lshlrev_b16 v115, 8, v4
-; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_3
-; GFX11-NEXT: ; %bb.1: ; %Flow
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB49_4
-; GFX11-NEXT: .LBB49_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB49_3: ; %cmp.false
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v36
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v38
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v31
-; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v35
-; GFX11-NEXT: v_or_b32_e32 v0, v0, v50
-; GFX11-NEXT: v_or_b32_e32 v1, v1, v51
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v34
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18
-; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-NEXT: v_or_b32_e32 v1, v3, v39
-; GFX11-NEXT: v_or_b32_e32 v3, v4, v54
-; GFX11-NEXT: v_or_b32_e32 v4, v5, v48
-; GFX11-NEXT: v_or_b32_e32 v5, v6, v55
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v32
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22
-; GFX11-NEXT: v_or_b32_e32 v2, v2, v49
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v53
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v52
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v17
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v83
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v81
-; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v4, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v6, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v8, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v10, v9, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24
-; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26
-; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28
-; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v71
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v65
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v80
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v68
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v87
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v84
-; GFX11-NEXT: v_or_b32_e32 v6, v6, v21
-; GFX11-NEXT: v_or_b32_e32 v7, v7, v19
-; GFX11-NEXT: v_or_b32_e32 v8, v8, v27
-; GFX11-NEXT: v_or_b32_e32 v9, v9, v23
-; GFX11-NEXT: v_or_b32_e32 v10, v10, v98
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v25
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v96
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v29
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v112
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v103
-; GFX11-NEXT: v_perm_b32 v6, v7, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v9, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v11, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v13, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v15, v14, 0x5040100
-; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v86
-; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v82
-; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v97
-; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v85
-; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v99
-; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v64
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v69
-; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v66
-; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v70
-; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v67
-; GFX11-NEXT: v_or_b32_e32 v11, v11, v101
-; GFX11-NEXT: v_or_b32_e32 v12, v12, v100
-; GFX11-NEXT: v_or_b32_e32 v13, v13, v114
-; GFX11-NEXT: v_or_b32_e32 v14, v14, v102
-; GFX11-NEXT: v_or_b32_e32 v15, v15, v117
-; GFX11-NEXT: v_or_b32_e32 v16, v16, v113
-; GFX11-NEXT: v_or_b32_e32 v17, v17, v116
-; GFX11-NEXT: v_or_b32_e32 v18, v18, v115
-; GFX11-NEXT: v_or_b32_e32 v19, v19, v118
-; GFX11-NEXT: v_or_b32_e32 v20, v20, v119
-; GFX11-NEXT: v_perm_b32 v11, v12, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v12, v14, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v16, v15, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v18, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v20, v19, 0x5040100
-; GFX11-NEXT: ; implicit-def: $vgpr36
-; GFX11-NEXT: ; implicit-def: $vgpr37
-; GFX11-NEXT: ; implicit-def: $vgpr33
-; GFX11-NEXT: ; implicit-def: $vgpr38
-; GFX11-NEXT: ; implicit-def: $vgpr31
-; GFX11-NEXT: ; implicit-def: $vgpr35
-; GFX11-NEXT: ; implicit-def: $vgpr34
-; GFX11-NEXT: ; implicit-def: $vgpr32
-; GFX11-NEXT: ; implicit-def: $vgpr16
-; GFX11-NEXT: ; implicit-def: $vgpr18
-; GFX11-NEXT: ; implicit-def: $vgpr20
-; GFX11-NEXT: ; implicit-def: $vgpr22
-; GFX11-NEXT: ; implicit-def: $vgpr24
-; GFX11-NEXT: ; implicit-def: $vgpr26
-; GFX11-NEXT: ; implicit-def: $vgpr28
-; GFX11-NEXT: ; implicit-def: $vgpr30
-; GFX11-NEXT: ; implicit-def: $vgpr71
-; GFX11-NEXT: ; implicit-def: $vgpr65
-; GFX11-NEXT: ; implicit-def: $vgpr80
-; GFX11-NEXT: ; implicit-def: $vgpr68
-; GFX11-NEXT: ; implicit-def: $vgpr87
-; GFX11-NEXT: ; implicit-def: $vgpr84
-; GFX11-NEXT: ; implicit-def: $vgpr86
-; GFX11-NEXT: ; implicit-def: $vgpr82
-; GFX11-NEXT: ; implicit-def: $vgpr97
-; GFX11-NEXT: ; implicit-def: $vgpr85
-; GFX11-NEXT: ; implicit-def: $vgpr99
-; GFX11-NEXT: ; implicit-def: $vgpr64
-; GFX11-NEXT: ; implicit-def: $vgpr69
-; GFX11-NEXT: ; implicit-def: $vgpr66
-; GFX11-NEXT: ; implicit-def: $vgpr70
-; GFX11-NEXT: ; implicit-def: $vgpr67
-; GFX11-NEXT: ; implicit-def: $vgpr50
-; GFX11-NEXT: ; implicit-def: $vgpr51
-; GFX11-NEXT: ; implicit-def: $vgpr49
-; GFX11-NEXT: ; implicit-def: $vgpr39
-; GFX11-NEXT: ; implicit-def: $vgpr54
-; GFX11-NEXT: ; implicit-def: $vgpr48
-; GFX11-NEXT: ; implicit-def: $vgpr55
-; GFX11-NEXT: ; implicit-def: $vgpr52
-; GFX11-NEXT: ; implicit-def: $vgpr53
-; GFX11-NEXT: ; implicit-def: $vgpr17
-; GFX11-NEXT: ; implicit-def: $vgpr83
-; GFX11-NEXT: ; implicit-def: $vgpr81
-; GFX11-NEXT: ; implicit-def: $vgpr21
-; GFX11-NEXT: ; implicit-def: $vgpr19
-; GFX11-NEXT: ; implicit-def: $vgpr27
-; GFX11-NEXT: ; implicit-def: $vgpr23
-; GFX11-NEXT: ; implicit-def: $vgpr98
-; GFX11-NEXT: ; implicit-def: $vgpr25
-; GFX11-NEXT: ; implicit-def: $vgpr96
-; GFX11-NEXT: ; implicit-def: $vgpr29
-; GFX11-NEXT: ; implicit-def: $vgpr112
-; GFX11-NEXT: ; implicit-def: $vgpr103
-; GFX11-NEXT: ; implicit-def: $vgpr101
-; GFX11-NEXT: ; implicit-def: $vgpr100
-; GFX11-NEXT: ; implicit-def: $vgpr114
-; GFX11-NEXT: ; implicit-def: $vgpr102
-; GFX11-NEXT: ; implicit-def: $vgpr117
-; GFX11-NEXT: ; implicit-def: $vgpr113
-; GFX11-NEXT: ; implicit-def: $vgpr116
-; GFX11-NEXT: ; implicit-def: $vgpr115
-; GFX11-NEXT: ; implicit-def: $vgpr118
-; GFX11-NEXT: ; implicit-def: $vgpr119
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB49_2
-; GFX11-NEXT: .LBB49_4: ; %cmp.true
-; GFX11-NEXT: v_add_nc_u16 v0, v70, 3
-; GFX11-NEXT: v_add_nc_u16 v1, v67, 3
-; GFX11-NEXT: v_add_nc_u16 v2, v69, 3
-; GFX11-NEXT: v_add_nc_u16 v3, v66, 3
-; GFX11-NEXT: v_add_nc_u16 v5, v64, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: s_waitcnt vmcnt(8)
-; GFX11-NEXT: v_add_nc_u16 v6, v97, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v118, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v119, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v99, 3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v0, v116, v2
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v5
-; GFX11-NEXT: v_or_b32_e32 v2, v115, v3
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v113, v1
-; GFX11-NEXT: v_add_nc_u16 v66, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v1, v114, v3
-; GFX11-NEXT: v_or_b32_e32 v4, v117, v4
-; GFX11-NEXT: v_add_nc_u16 v2, v85, 3
-; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(6)
-; GFX11-NEXT: v_add_nc_u16 v0, v86, 3
-; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v82, 3
-; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v4
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: s_waitcnt vmcnt(4)
-; GFX11-NEXT: v_add_nc_u16 v3, v87, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v84, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v102, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v101, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v100, v1
-; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v112, v3
-; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v103, v4
-; GFX11-NEXT: v_add_nc_u16 v70, 0x300, v1
-; GFX11-NEXT: s_waitcnt vmcnt(2)
-; GFX11-NEXT: v_add_nc_u16 v1, v80, 3
-; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v68, 3
-; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v0
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_add_nc_u16 v0, v71, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v3, v65, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v4, v28, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_or_b32_e32 v1, v96, v1
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v2, v29, v2
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v0, v98, v0
-; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v25, v3
-; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v27, v4
-; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v30, 3
-; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v24, 3
-; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v26, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v3, v20, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_add_nc_u16 v4, v22, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_or_b32_e32 v0, v23, v0
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v1, v21, v1
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v2, v19, v2
-; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v83, v3
-; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v1
-; GFX11-NEXT: v_or_b32_e32 v1, v81, v4
-; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2
-; GFX11-NEXT: v_add_nc_u16 v2, v16, 3
-; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v0, v34, 3
-; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v31, 3
-; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-NEXT: v_add_nc_u16 v3, v18, 3
-; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-NEXT: v_add_nc_u16 v4, v32, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_or_b32_e32 v2, v53, v2
-; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-NEXT: v_or_b32_e32 v0, v55, v0
-; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-NEXT: v_or_b32_e32 v1, v54, v1
-; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v2
-; GFX11-NEXT: v_or_b32_e32 v2, v17, v3
-; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v52, v4
-; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v1, v33, 3
-; GFX11-NEXT: v_add_nc_u16 v17, v36, 3
-; GFX11-NEXT: v_add_nc_u16 v21, v37, 3
-; GFX11-NEXT: v_add_nc_u16 v22, v38, 3
-; GFX11-NEXT: v_add_nc_u16 v23, v35, 3
-; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22
-; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23
-; GFX11-NEXT: v_or_b32_e32 v1, v49, v1
-; GFX11-NEXT: v_or_b32_e32 v17, v50, v17
-; GFX11-NEXT: v_or_b32_e32 v21, v51, v21
-; GFX11-NEXT: v_or_b32_e32 v22, v39, v22
-; GFX11-NEXT: v_or_b32_e32 v23, v48, v23
-; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1
-; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17
-; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21
-; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22
-; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23
-; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0
-; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2
-; GFX11-NEXT: v_perm_b32 v0, v21, v17, 0x5040100
-; GFX11-NEXT: v_perm_b32 v1, v22, v1, 0x5040100
-; GFX11-NEXT: v_perm_b32 v2, v23, v4, 0x5040100
-; GFX11-NEXT: v_perm_b32 v3, v24, v3, 0x5040100
-; GFX11-NEXT: v_perm_b32 v4, v26, v18, 0x5040100
-; GFX11-NEXT: v_perm_b32 v5, v16, v5, 0x5040100
-; GFX11-NEXT: v_perm_b32 v6, v20, v6, 0x5040100
-; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100
-; GFX11-NEXT: v_perm_b32 v8, v27, v8, 0x5040100
-; GFX11-NEXT: v_perm_b32 v9, v25, v9, 0x5040100
-; GFX11-NEXT: v_perm_b32 v10, v68, v10, 0x5040100
-; GFX11-NEXT: v_perm_b32 v11, v70, v11, 0x5040100
-; GFX11-NEXT: v_perm_b32 v12, v69, v12, 0x5040100
-; GFX11-NEXT: v_perm_b32 v13, v67, v13, 0x5040100
-; GFX11-NEXT: v_perm_b32 v14, v66, v14, 0x5040100
-; GFX11-NEXT: v_perm_b32 v15, v64, v15, 0x5040100
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = add <64 x i8> %a, splat (i8 3)
- %a2 = bitcast <64 x i8> %a1 to <32 x i16>
- br label %end
-
-cmp.false:
- %a3 = bitcast <64 x i8> %a to <32 x i16>
- br label %end
-
-end:
- %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x i16> %phi
-}
-
-define <32 x bfloat> @bitcast_v32f16_to_v32bf16(<32 x half> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32f16_to_v32bf16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v0
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v40, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v41, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v42, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v43, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v44, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v45, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v46, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v47, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v56, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v57, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v58, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v59, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v60, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v61, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v62, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v63
-; GCN-NEXT: v_cvt_f16_f32_e32 v63, v31
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v55
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v40
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v41
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v42
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v43
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v45
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v46
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v47
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v56
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v57
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v58
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v59
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v60
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v61
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v62
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v63
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB50_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB50_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v59
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v58
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v57
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v56
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v47
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v46
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v45
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v44
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v43
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v42
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v41
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v40
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v32
-; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31
-; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30
-; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29
-; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28
-; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27
-; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26
-; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25
-; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24
-; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23
-; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22
-; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21
-; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20
-; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19
-; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18
-; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17
-; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16
-; GCN-NEXT: v_cvt_f16_f32_e32 v32, v15
-; GCN-NEXT: v_cvt_f16_f32_e32 v33, v14
-; GCN-NEXT: v_cvt_f16_f32_e32 v34, v13
-; GCN-NEXT: v_cvt_f16_f32_e32 v35, v12
-; GCN-NEXT: v_cvt_f16_f32_e32 v36, v11
-; GCN-NEXT: v_cvt_f16_f32_e32 v37, v10
-; GCN-NEXT: v_cvt_f16_f32_e32 v38, v9
-; GCN-NEXT: v_cvt_f16_f32_e32 v39, v8
-; GCN-NEXT: v_cvt_f16_f32_e32 v48, v7
-; GCN-NEXT: v_cvt_f16_f32_e32 v49, v6
-; GCN-NEXT: v_cvt_f16_f32_e32 v50, v5
-; GCN-NEXT: v_cvt_f16_f32_e32 v51, v4
-; GCN-NEXT: v_cvt_f16_f32_e32 v52, v3
-; GCN-NEXT: v_cvt_f16_f32_e32 v53, v2
-; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1
-; GCN-NEXT: v_cvt_f16_f32_e32 v55, v0
-; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v31
-; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
-; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29
-; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v28
-; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v27
-; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v25
-; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v24
-; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v23
-; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v22
-; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v20
-; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v19
-; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v18
-; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v17
-; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v16
-; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v32
-; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v33
-; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v34
-; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v35
-; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v36
-; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v37
-; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v38
-; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v39
-; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v48
-; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v49
-; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v50
-; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v51
-; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v52
-; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v53
-; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v54
-; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v55
-; GCN-NEXT: .LBB50_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32f16_to_v32bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB50_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_mov_b32_e32 v17, 0x200
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v15
-; VI-NEXT: v_add_f16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v15, v19, v15
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v14
-; VI-NEXT: v_add_f16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v14, v19, v14
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v13
-; VI-NEXT: v_add_f16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v13, v19, v13
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v12
-; VI-NEXT: v_add_f16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v12, v19, v12
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v11
-; VI-NEXT: v_add_f16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v11, v19, v11
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v10
-; VI-NEXT: v_add_f16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v10, v19, v10
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v9
-; VI-NEXT: v_add_f16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v9, v19, v9
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v8
-; VI-NEXT: v_add_f16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v8, v19, v8
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v7
-; VI-NEXT: v_add_f16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v7, v19, v7
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v6
-; VI-NEXT: v_add_f16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v6, v19, v6
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v5
-; VI-NEXT: v_add_f16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v5, v19, v5
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v4
-; VI-NEXT: v_add_f16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v16, 0x200, v0
-; VI-NEXT: v_add_f16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v18, 0x200, v1
-; VI-NEXT: v_add_f16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v4, v19, v4
-; VI-NEXT: v_add_f16_e32 v19, 0x200, v2
-; VI-NEXT: v_add_f16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_sdwa v17, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
-; VI-NEXT: v_or_b32_e32 v3, v3, v17
-; VI-NEXT: v_or_b32_e32 v2, v19, v2
-; VI-NEXT: v_or_b32_e32 v1, v18, v1
-; VI-NEXT: v_or_b32_e32 v0, v16, v0
-; VI-NEXT: .LBB50_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32f16_to_v32bf16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB50_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: s_movk_i32 s6, 0x200
-; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB50_2: ; %end
-; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: bitcast_v32f16_to_v32bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-NEXT: s_cbranch_execz .LBB50_2
-; GFX11-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-NEXT: .LBB50_2: ; %end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
- %cmp = icmp eq i32 %b, 0
- br i1 %cmp, label %cmp.true, label %cmp.false
-
-cmp.true:
- %a1 = fadd <32 x half> %a, splat (half 0xH0200)
- %a2 = bitcast <32 x half> %a1 to <32 x bfloat>
- br label %end
-
-cmp.false:
- %a3 = bitcast <32 x half> %a to <32 x bfloat>
- br label %end
-
-end:
- %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
- ret <32 x bfloat> %phi
-}
-
-define <32 x half> @bitcast_v32bf16_to_v32f16(<32 x bfloat> %a, i32 %b) {
-; GCN-LABEL: bitcast_v32bf16_to_v32f16:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; GCN-NEXT: s_waitcnt expcnt(0)
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:4
-; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0
-; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v1
-; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v2
-; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v3
-; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v4
-; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v5
-; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v6
-; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v7
-; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v8
-; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v9
-; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v10
-; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v11
-; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v12
-; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13
-; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v14
-; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v15
-; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v16
-; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v17
-; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v18
-; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v19
-; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v20
-; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v21
-; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v22
-; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v23
-; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v24
-; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v25
-; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v26
-; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v27
-; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v28
-; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v29
-; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v30
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v63
-; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v31
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
-; GCN-NEXT: ; implicit-def: $vgpr3
-; GCN-NEXT: ; implicit-def: $vgpr4
-; GCN-NEXT: ; implicit-def: $vgpr5
-; GCN-NEXT: ; implicit-def: $vgpr6
-; GCN-NEXT: ; implicit-def: $vgpr7
-; GCN-NEXT: ; implicit-def: $vgpr8
-; GCN-NEXT: ; implicit-def: $vgpr9
-; GCN-NEXT: ; implicit-def: $vgpr10
-; GCN-NEXT: ; implicit-def: $vgpr11
-; GCN-NEXT: ; implicit-def: $vgpr12
-; GCN-NEXT: ; implicit-def: $vgpr13
-; GCN-NEXT: ; implicit-def: $vgpr14
-; GCN-NEXT: ; implicit-def: $vgpr15
-; GCN-NEXT: ; implicit-def: $vgpr16
-; GCN-NEXT: ; implicit-def: $vgpr17
-; GCN-NEXT: ; implicit-def: $vgpr18
-; GCN-NEXT: ; implicit-def: $vgpr19
-; GCN-NEXT: ; implicit-def: $vgpr20
-; GCN-NEXT: ; implicit-def: $vgpr21
-; GCN-NEXT: ; implicit-def: $vgpr22
-; GCN-NEXT: ; implicit-def: $vgpr23
-; GCN-NEXT: ; implicit-def: $vgpr24
-; GCN-NEXT: ; implicit-def: $vgpr25
-; GCN-NEXT: ; implicit-def: $vgpr26
-; GCN-NEXT: ; implicit-def: $vgpr27
-; GCN-NEXT: ; implicit-def: $vgpr28
-; GCN-NEXT: ; implicit-def: $vgpr29
-; GCN-NEXT: ; implicit-def: $vgpr30
-; GCN-NEXT: ; implicit-def: $vgpr31
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_2
-; GCN-NEXT: ; %bb.1: ; %cmp.false
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v32
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v33
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v34
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v36
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v37
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v38
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v39
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v48
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v49
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v50
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v51
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v52
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v53
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v54
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v55
-; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v40
-; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v41
-; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v42
-; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v43
-; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v44
-; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v45
-; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v46
-; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v47
-; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v56
-; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v57
-; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v58
-; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v59
-; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v60
-; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v61
-; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v62
-; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v63
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31
-; GCN-NEXT: ; implicit-def: $vgpr32
-; GCN-NEXT: ; implicit-def: $vgpr33
-; GCN-NEXT: ; implicit-def: $vgpr34
-; GCN-NEXT: ; implicit-def: $vgpr35
-; GCN-NEXT: ; implicit-def: $vgpr36
-; GCN-NEXT: ; implicit-def: $vgpr37
-; GCN-NEXT: ; implicit-def: $vgpr38
-; GCN-NEXT: ; implicit-def: $vgpr39
-; GCN-NEXT: ; implicit-def: $vgpr48
-; GCN-NEXT: ; implicit-def: $vgpr49
-; GCN-NEXT: ; implicit-def: $vgpr50
-; GCN-NEXT: ; implicit-def: $vgpr51
-; GCN-NEXT: ; implicit-def: $vgpr52
-; GCN-NEXT: ; implicit-def: $vgpr53
-; GCN-NEXT: ; implicit-def: $vgpr54
-; GCN-NEXT: ; implicit-def: $vgpr55
-; GCN-NEXT: ; implicit-def: $vgpr40
-; GCN-NEXT: ; implicit-def: $vgpr41
-; GCN-NEXT: ; implicit-def: $vgpr42
-; GCN-NEXT: ; implicit-def: $vgpr43
-; GCN-NEXT: ; implicit-def: $vgpr44
-; GCN-NEXT: ; implicit-def: $vgpr45
-; GCN-NEXT: ; implicit-def: $vgpr46
-; GCN-NEXT: ; implicit-def: $vgpr47
-; GCN-NEXT: ; implicit-def: $vgpr56
-; GCN-NEXT: ; implicit-def: $vgpr57
-; GCN-NEXT: ; implicit-def: $vgpr58
-; GCN-NEXT: ; implicit-def: $vgpr59
-; GCN-NEXT: ; implicit-def: $vgpr60
-; GCN-NEXT: ; implicit-def: $vgpr61
-; GCN-NEXT: ; implicit-def: $vgpr62
-; GCN-NEXT: ; implicit-def: $vgpr63
-; GCN-NEXT: .LBB51_2: ; %Flow
-; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GCN-NEXT: s_cbranch_execz .LBB51_4
-; GCN-NEXT: ; %bb.3: ; %cmp.true
-; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v63
-; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62
-; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61
-; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60
-; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v59
-; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v58
-; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v57
-; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v56
-; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v47
-; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v46
-; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v45
-; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v44
-; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v43
-; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v42
-; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v41
-; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v40
-; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v55
-; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v54
-; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v53
-; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v52
-; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v51
-; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50
-; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v49
-; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v48
-; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v39
-; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v38
-; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v37
-; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v36
-; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v35
-; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v34
-; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v33
-; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v32
-; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0
-; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1
-; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2
-; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4
-; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5
-; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6
-; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7
-; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v9
-; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v10
-; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v11
-; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v12
-; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v13
-; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v14
-; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v15
-; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v16
-; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v17
-; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v18
-; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v19
-; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v20
-; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v21
-; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v22
-; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v23
-; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v24
-; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v25
-; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v26
-; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v27
-; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v28
-; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v29
-; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v30
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v31
-; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
-; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6
-; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7
-; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8
-; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9
-; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11
-; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12
-; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13
-; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14
-; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15
-; GCN-NEXT: v_cvt_f32_f16_e32 v16, v55
-; GCN-NEXT: v_cvt_f32_f16_e32 v17, v54
-; GCN-NEXT: v_cvt_f32_f16_e32 v18, v53
-; GCN-NEXT: v_cvt_f32_f16_e32 v19, v52
-; GCN-NEXT: v_cvt_f32_f16_e32 v20, v51
-; GCN-NEXT: v_cvt_f32_f16_e32 v21, v50
-; GCN-NEXT: v_cvt_f32_f16_e32 v22, v49
-; GCN-NEXT: v_cvt_f32_f16_e32 v23, v48
-; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39
-; GCN-NEXT: v_cvt_f32_f16_e32 v25, v38
-; GCN-NEXT: v_cvt_f32_f16_e32 v26, v37
-; GCN-NEXT: v_cvt_f32_f16_e32 v27, v36
-; GCN-NEXT: v_cvt_f32_f16_e32 v28, v35
-; GCN-NEXT: v_cvt_f32_f16_e32 v29, v34
-; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33
-; GCN-NEXT: v_cvt_f32_f16_e32 v31, v32
-; GCN-NEXT: .LBB51_4: ; %end
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v32bf16_to_v32f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB51_2
-; VI-NEXT: ; %bb.1: ; %cmp.true
-; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; VI-NEXT: v_bfe_u32 v17, v16, 16, 1
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17
-; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; VI-NEXT: v_bfe_u32 v17, v0, 16, 1
-; VI-NEXT: s_movk_i32 s6, 0x7fff
-; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0
-; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17
-; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; VI-NEXT: v_bfe_u32 v18, v17, 16, 1
-; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17
-; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc
-; VI-NEXT: v_bfe_u32 v18, v1, 16, 1
-; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v1
-; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18
-; VI-NEXT: v_or_b32_e32 v19, 0x400000, v1
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; VI-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; VI-NEXT: v_bfe_u32 v19, v18, 16, 1
-; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v18
-; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19
-; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; VI-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; VI-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc
-; VI-NEXT: v_bfe_u32 v19, v2, 16, 1
-; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v2
-; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19
-; VI-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; VI-NEXT: v_bfe_u32 v20, v19, 16, 1
-; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v19
-; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20
-; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; VI-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; VI-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc
-; VI-NEXT: v_bfe_u32 v20, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v3
-; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20
-; VI-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v4
-; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; VI-NEXT: v_bfe_u32 v21, v20, 16, 1
-; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v20
-; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21
-; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; VI-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; VI-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc
-; VI-NEXT: v_bfe_u32 v21, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v4
-; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21
-; VI-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; VI-NEXT: v_bfe_u32 v22, v21, 16, 1
-; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v21
-; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; VI-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; VI-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc
-; VI-NEXT: v_bfe_u32 v22, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v5
-; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22
-; VI-NEXT: v_or_b32_e32 v23, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v6
-; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; VI-NEXT: v_bfe_u32 v23, v22, 16, 1
-; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v22
-; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23
-; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; VI-NEXT: v_or_b32_e32 v24, 0x400000, v22
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; VI-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc
-; VI-NEXT: v_bfe_u32 v23, v6, 16, 1
-; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v6
-; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23
-; VI-NEXT: v_or_b32_e32 v24, 0x400000, v6
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; VI-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v7
-; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; VI-NEXT: v_bfe_u32 v24, v23, 16, 1
-; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v23
-; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24
-; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; VI-NEXT: v_or_b32_e32 v25, 0x400000, v23
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; VI-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc
-; VI-NEXT: v_bfe_u32 v24, v7, 16, 1
-; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v7
-; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24
-; VI-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; VI-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v8
-; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; VI-NEXT: v_bfe_u32 v25, v24, 16, 1
-; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v24
-; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25
-; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; VI-NEXT: v_or_b32_e32 v26, 0x400000, v24
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; VI-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc
-; VI-NEXT: v_bfe_u32 v25, v8, 16, 1
-; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v8
-; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25
-; VI-NEXT: v_or_b32_e32 v26, 0x400000, v8
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; VI-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v9
-; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
-; VI-NEXT: v_bfe_u32 v26, v25, 16, 1
-; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v25
-; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26
-; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
-; VI-NEXT: v_or_b32_e32 v27, 0x400000, v25
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
-; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; VI-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc
-; VI-NEXT: v_bfe_u32 v26, v9, 16, 1
-; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v9
-; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26
-; VI-NEXT: v_or_b32_e32 v27, 0x400000, v9
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
-; VI-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v10
-; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
-; VI-NEXT: v_bfe_u32 v27, v26, 16, 1
-; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v26
-; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27
-; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; VI-NEXT: v_or_b32_e32 v28, 0x400000, v26
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
-; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; VI-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc
-; VI-NEXT: v_bfe_u32 v27, v10, 16, 1
-; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v10
-; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27
-; VI-NEXT: v_or_b32_e32 v28, 0x400000, v10
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
-; VI-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v11
-; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; VI-NEXT: v_bfe_u32 v28, v27, 16, 1
-; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v27
-; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28
-; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
-; VI-NEXT: v_or_b32_e32 v29, 0x400000, v27
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
-; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; VI-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc
-; VI-NEXT: v_bfe_u32 v28, v11, 16, 1
-; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v11
-; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28
-; VI-NEXT: v_or_b32_e32 v29, 0x400000, v11
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
-; VI-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v12
-; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
-; VI-NEXT: v_bfe_u32 v29, v28, 16, 1
-; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v28
-; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29
-; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; VI-NEXT: v_or_b32_e32 v30, 0x400000, v28
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
-; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; VI-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc
-; VI-NEXT: v_bfe_u32 v29, v12, 16, 1
-; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v12
-; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29
-; VI-NEXT: v_or_b32_e32 v30, 0x400000, v12
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
-; VI-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v13
-; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
-; VI-NEXT: v_bfe_u32 v30, v29, 16, 1
-; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v29
-; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30
-; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
-; VI-NEXT: v_or_b32_e32 v31, 0x400000, v29
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
-; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; VI-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc
-; VI-NEXT: v_bfe_u32 v30, v13, 16, 1
-; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v13
-; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30
-; VI-NEXT: v_or_b32_e32 v31, 0x400000, v13
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
-; VI-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v14
-; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
-; VI-NEXT: v_bfe_u32 v31, v30, 16, 1
-; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v30
-; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31
-; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
-; VI-NEXT: v_or_b32_e32 v32, 0x400000, v30
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
-; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; VI-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc
-; VI-NEXT: v_bfe_u32 v31, v14, 16, 1
-; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v14
-; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31
-; VI-NEXT: v_or_b32_e32 v32, 0x400000, v14
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
-; VI-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc
-; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v15
-; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; VI-NEXT: v_bfe_u32 v32, v31, 16, 1
-; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v31
-; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32
-; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; VI-NEXT: v_or_b32_e32 v33, 0x400000, v31
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
-; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; VI-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
-; VI-NEXT: v_bfe_u32 v32, v15, 16, 1
-; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v15
-; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32
-; VI-NEXT: v_or_b32_e32 v33, 0x400000, v15
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
-; VI-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc
-; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
-; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14
-; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
-; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12
-; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
-; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10
-; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
-; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8
-; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
-; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; VI-NEXT: v_alignbit_b32 v15, v15, v31, 16
-; VI-NEXT: v_alignbit_b32 v14, v14, v30, 16
-; VI-NEXT: v_alignbit_b32 v13, v13, v29, 16
-; VI-NEXT: v_alignbit_b32 v12, v12, v28, 16
-; VI-NEXT: v_alignbit_b32 v11, v11, v27, 16
-; VI-NEXT: v_alignbit_b32 v10, v10, v26, 16
-; VI-NEXT: v_alignbit_b32 v9, v9, v25, 16
-; VI-NEXT: v_alignbit_b32 v8, v8, v24, 16
-; VI-NEXT: v_alignbit_b32 v7, v7, v23, 16
-; VI-NEXT: v_alignbit_b32 v6, v6, v22, 16
-; VI-NEXT: v_alignbit_b32 v5, v5, v21, 16
-; VI-NEXT: v_alignbit_b32 v4, v4, v20, 16
-; VI-NEXT: v_alignbit_b32 v3, v3, v19, 16
-; VI-NEXT: v_alignbit_b32 v2, v2, v18, 16
-; VI-NEXT: v_alignbit_b32 v1, v1, v17, 16
-; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16
-; VI-NEXT: .LBB51_2: ; %end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v32bf16_to_v32f16:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
-; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB51_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.true
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0
-; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
-; GFX9-NEXT: s_movk_i32 s6, 0x7fff
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
-; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
-; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1
-; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6
-; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
-; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
-; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX9-NEXT: v_bfe_u32 v18, v17, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX9-NEXT: v_add3_u32 v18, v18, v17, s6
-; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v17
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
-; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc
-; GFX9-NEXT: v_bfe_u32 v18, v1, 16, 1
-; GFX9-NEXT: v_add3_u32 v18, v18, v1, s6
-; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v1
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX9-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v2
-; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX9-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX9-NEXT: v_add3_u32 v19, v19, v18, s6
-; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v18
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
-; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc
-; GFX9-NEXT: v_bfe_u32 v19, v2, 16, 1
-; GFX9-NEXT: v_add3_u32 v19, v19, v2, s6
-; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
-; GFX9-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v3
-; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX9-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX9-NEXT: v_add3_u32 v20, v20, v19, s6
-; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
-; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc
-; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6
-; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v4
-; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX9-NEXT: v_bfe_u32 v21, v20, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX9-NEXT: v_add3_u32 v21, v21, v20, s6
-; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
-; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc
-; GFX9-NEXT: v_bfe_u32 v21, v4, 16, 1
-; GFX9-NEXT: v_add3_u32 v21, v21, v4, s6
-; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX9-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v5
-; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX9-NEXT: v_bfe_u32 v22, v21, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
-; GFX9-NEXT: v_add3_u32 v22, v22, v21, s6
-; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
-; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc
-; GFX9-NEXT: v_bfe_u32 v22, v5, 16, 1
-; GFX9-NEXT: v_add3_u32 v22, v22, v5, s6
-; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v5
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX9-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v6
-; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX9-NEXT: v_bfe_u32 v23, v22, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
-; GFX9-NEXT: v_add3_u32 v23, v23, v22, s6
-; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v22
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
-; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc
-; GFX9-NEXT: v_bfe_u32 v23, v6, 16, 1
-; GFX9-NEXT: v_add3_u32 v23, v23, v6, s6
-; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v6
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX9-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v7
-; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
-; GFX9-NEXT: v_bfe_u32 v24, v23, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
-; GFX9-NEXT: v_add3_u32 v24, v24, v23, s6
-; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v23
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
-; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc
-; GFX9-NEXT: v_bfe_u32 v24, v7, 16, 1
-; GFX9-NEXT: v_add3_u32 v24, v24, v7, s6
-; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
-; GFX9-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc
-; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v8
-; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX9-NEXT: v_bfe_u32 v25, v24, 16, 1
-; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
-; GFX9-NEXT: v_add3_u32 v25, v25, v24, s6
-; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v24
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
-; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX9-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc
-; GFX9-NEXT: v_bfe_u32 v25, v8, 16, 1
-; GFX9-NEXT: v_add3_u32 v25, v25, v8, s6
-; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v8
-; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
-; GFX9-NEXT: v_cndmask_